From dfa1ae9a99acbd8b74e5cd13b61baf569eb54348 Mon Sep 17 00:00:00 2001 From: Cursor Agent Date: Mon, 21 Sep 2026 02:10:32 +0000 Subject: [PATCH 1/2] Fold hourly 1946 HIGH MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Novel HIGH catalog/class-table fold: X-sentiment does not execute, heyjunpenn 485 catalog, jev-arena AI-reviewed labels, one-trial robot, Qwen3.8 10.59x uncalibrated, Spanish acento audit. Light densify alexwestco/llm-to-jev description rewrite (SHA unchanged). Skip Open-Jev #53 and #54 three. IDs: notes.md §131 / items 505-520 / #113. Co-authored-by: Basit Mustafa <24601@users.noreply.github.com> --- .agents/skills/augustus/SKILL.md | 5 +- .../references/agent-self-assessment.md | 2 + .../augustus/references/applied-mappings.md | 2 + .../references/composition-algebra.md | 72 + .agents/skills/augustus/references/faq.md | 2 + .../augustus/references/formal-methods.md | 2 + .../augustus/references/formal-semi-formal.md | 2 + .../augustus/references/judgment-class.md | 2 + .../skills/augustus/references/mappings.md | 2 + .../augustus/references/mental-models.md | 13 + .../augustus/references/methods-catalog.md | 2 + .../augustus/references/mixed-architecture.md | 2 + .../augustus/references/question-design.md | 2 + .../augustus/references/toolbox-mapping.md | 2 + .../skills/augustus/references/validation.md | 2 + .../augustus/scripts/evaluate_decisions.py | 61 + .../augustus/scripts/uniqueness_gate.py | 95 +- CHANGELOG.md | 30 + CONTRIBUTING.md | 2 +- README.md | 2 + docs/_includes/recipes.html | 42 + docs/ecosystem.md | 2 + research/archive/findings.md | 31 +- .../hourly/2026-09-21T01/augustus_items.json | 668 +++++++ .../archive/hourly/2026-09-21T01/github.json | 1768 +++++++++++++++++ .../2026-09-21T01/novel_high_this_run.json | 1746 ++++++++++++++++ .../hourly/2026-09-21T01/prompt_Augustus.md | 102 + .../archive/hourly/2026-09-21T01/readmes.json | 281 +++ .../2026-09-21T01/revisit_high_this_run.json | 24 + .../hourly/2026-09-21T01/run_digest.json | 11 + research/changelog-hourly.md | 12 + research/notes.md | 325 +++ research/refresh-log.md | 18 +- research/revisit_fingerprints.json | 158 +- 34 files changed, 5481 insertions(+), 11 deletions(-) create mode 100644 research/archive/hourly/2026-09-21T01/augustus_items.json create mode 100644 research/archive/hourly/2026-09-21T01/github.json create mode 100644 research/archive/hourly/2026-09-21T01/novel_high_this_run.json create mode 100644 research/archive/hourly/2026-09-21T01/prompt_Augustus.md create mode 100644 research/archive/hourly/2026-09-21T01/readmes.json create mode 100644 research/archive/hourly/2026-09-21T01/revisit_high_this_run.json create mode 100644 research/archive/hourly/2026-09-21T01/run_digest.json diff --git a/.agents/skills/augustus/SKILL.md b/.agents/skills/augustus/SKILL.md index 9c5ea9b..1c84245 100644 --- a/.agents/skills/augustus/SKILL.md +++ b/.agents/skills/augustus/SKILL.md @@ -54,7 +54,7 @@ classical method you already trust, substitute it, classify the win "paraphrase brittleness", "allowlist then judge", "TOCTOU-of-Noul", "Jev inside the database / sqlite-jev", "Jev picks bitrate / join order / the model", "wait for Archer", "lint the request / missing - other", "training confronts Choice other / none-of-the-above", "soft AGENTS.md rules vs the linter", "screenshot Choice / omni System One", "extractive quotes / pointer not generator", "compaction summarize vs pointer", "encoder vs Jev compaction backend", "shadow-mode compaction rollout", "CI flaky-vs-real merge gate", "fail-open VOI wake/resume", "claim vs session evidence", "S1 indexer escalate-S2", "Harbor on/off routing", "fail-open vs fail-closed wake vs CI gate", "encoder vs Jev computer-use backend", "hybrid local decide + remote fill", "DONE vs verified success", "stdout prune vs session compaction", "OpenCode jev-pruner vs Claude jev-pruner", "zen-chat vs jev-zen Noul", "hard envelope then Noul prune", "Cua-S1 vs TypeSafe Jev", "plan vs execute dry-run", "specialist computer-use vs general agent", "local drop-in vs stub scorer", "route vs memory", "when does it hold / extractable from state", "decision model vs constrained LLM", "dual-process S1/S2", "combinatorial grid vs extractive", "uncalibrated local likelihoods", "decision-native RAG", "classify-first / read selectively", "living applied-mappings atlas / class patterns", "silence as safer / draft-gate heartbeat", "robotics text-state vs pixels", "verbatim ledger vs summary", "judgment as language primitive", "Stagehand extract pick-and-copy", "harness observe-score-act vs demo loop", "public judgment wall / six parallel questions", "meaning-search without embeddings", "attention ≠ correctness", "skills→oxlint / AST prove ∩ remainder", "session-sticky first-prompt routing", "measured RAG rerank vs generative rerank", "capability kernel / secrets never in the agent", "Jev is SENSOR not policy", "type-safe ≠ correct", "typed control plane around DSPy", "native vs verbalized confidence", "engine owns truth / Jev owns judgment", "human-confirmed kill gate", "train specialist vs few-shot hosted", "decide→policy→LLM leftover", "Noul 0.5 cannot-tell never rounded", "calibration ≠ sortable / ORDER BY", "pairwise inversion / Score ordinality / two-decimal ties", "wire-compat GLiFormer /v1/systemone", "class-backend economics", "loopback gateway hosted + local", "do not distill Jev as teacher", "active-learning triage", "evidence-packet explorer", "meaning-grep AND/OR/NOT", "closed-vote-only / no planner LLM", "Jev vs PCD Harbor", "PCD O(1) ≠ Noul", "host-owned handlers × System One", "OMP/pi fail-open gate", "permission vs probability / operator owns thresholds", "judgment ≠ permission / Jev never grants access", "eval integrity / instrument not score", "constrained optimizer + S1 features / never sole hot-path gate", "privilege ≠ verdict / effect contracts not tokens", "attention filter / VOI for human review / never blocks / never green unless sure", "measurement owns endorsement / evidence-gated question packs", "Jev supplies evidence / code owns authority", "ranking ≠ calibration / never hard-threshold raw p as frequency", "hot-click CU / indexed element table", "Jev judges relevance / code decides structure", "local rules first then remainder / never auto-train on own hides", "combinators / System One as control plane", "receipts not leaderboard / type-safe ≠ correct jaggedness", "VOI over skill library / skillranker abstention", "OOD calibration / AUC ≠ ECE", "Jev vs thinking-budget small models", "turnstile / replayable evidence≠authority", "MLX one-pass schema→JSON / Apple Silicon replica economics", "memory leases ended by new evidence", "never confidently wrong / TLA+ compose / escalate instead of hard-gate", "no seal no advance / coverage ledger / mint ≠ product brain", "skill-broker sibling / judgment ≠ permission", "sureness bands / max_prob is generous", "JevBench / calibration not in Main Score", "CI typed gate before expensive review", "Codex MCP host adapter", "judgment as attention redirect / jev-preflight", "compress-before-first-send / dizk jev-lens", "tools≠use / SessionStart over hoping", "observational memory / pi-om keep-kind", "open-Jev class / openvons / JevPick", "physical-world System One / HA-Jev / not for locks", "judgment outside the store / jevql", "landed-script trust / headless≠auto-approve", "digital-design combinators / extended five", "VOI cache admission / same-intent skip LLM", "BM25 vs Jev skill routing Harbor harness", "zeroshot vs BERT / contamination DiD", "typed escalate continue abort baton / inverted loop", "worth-your-attention VOI / ThinkyMiner Winnow", "Jev WHETHER Python HOW LLM WHAT", "conflict vs ignorance / named Choice escape", "Playwright executes Jev chooses", "OpenJev /v1/decide not drop-in", "SemIf wire-compat runoff; SemIf rename densify / MLX backend / 5.21× systems≠semantic / Softmax ≠ Noul (`notes.md` §117)", "decision-as-memory flywheel", "record/replay CI / jevassert", "failure-finding arena / jevarena ≠ jev-arena", "BBQ not a bias cert", "decider≠executor", "sentence-as-rule lint / jevlint", "sentence-as-rule lint / jev-lint is jevlint rename", "VOI hunk prune", "whole-repo intent VERIFIED/VIOLATION/UNKNOWN", "GLiNER2 spec ≠ replica", "open replica substrates / grande / laya-jolt / JEV-CPU", "ONNX local-jev not equivalent", "persist constraints across compaction / pi-heed", "calibration+cost first-class gates", "Harbor-shaped Jev vs SGR LLM-as-judge / jev-judge-bench ≠ jevarena ≠ jevbench", "hand no-text steps / jev-use / Vercel drops confidence", "Pi System-One control plane / pi-jev-control", "never free-generates / jev-gpt tree of Choices", "OpenRouter recipe atlas / samples not benches", "personal history feed / jevfeed / no social graph", "competing NAR claims / dual-channel ECE / openJev-verdict ≠ OpenJev", "empty compaction-proxy skip / IPECTER", "throughput ≠ latency / like-for-like ECE", "1-token logprob endpoint ≠ Noul / coverage ≠ correctness", "open replica engine / jevinf", "unofficial Elixir SDK ≠ OTP peer", "jevex n=16 files-to-read VOI", "commit pre-review attention≠verdict / middle band", "Hermes plugin is Agnes not TypeSafe", "pi-jev-compact ≠ pi-jev-compaction", "empty Codex-proxy skip / IPECTER runway", "decision-native inbox / mailordinal", "unofficial jev-cli not ready / ≠ jevql", "laya-multilingual / English checkpoint confident-wrong OOD", "schema-scorer peaked ranking ≠ calibration", "HF 401 / GitHub 404 Hub-only", "productized System One HTTP / classifier.dev", "escalate-under-threshold / smart tier / multi-label ignores", "silent FALLBACK / granite 0.546 vs advertised 0.800", "vs_jev tracked JSON / read eval/README", "choxos/jev-reviewer ≠ egma-ai / systematic-review pointer", "two-pass Choice+Noul evidence extraction", "not-found is an answer", "human check as productized judgment", "githubnext/localjev ≠ kunchenguid/local-jev", "wire-compat ≠ logit-equiv / prompted JSON ≠ structured read", "self-reported probs / entropy confidence", "GitHub Next local /v1/systemone", "LM Studio runner gap / structured-read primitives", "NandhaKishorM/laya packaging ≠ Hub-only / Router script-before-p", "post-T ECE ≠ raw ECE / Banking77 token-budget", "0.85 still soft / not TypeSafe drop-in", "external census ≠ scored bake-off", "GLiNER2+routers class-boundary", "incomplete openjev census vs watch", "Harbor honesty watch / silent fallback", "JevBench v1.2 geometric mean / cal ON rank / weight sensitivity", "option-order 72→21 / instruction models in the class table", "self-host latency ×2 assumption / est. costs", "Laya absent is a gap not a named exclusion", "Qwen3.8 27B ≠ Archer", "hourly already-folded watch / apply-the-five / skip thin noise", "hard-gate Noul as PR/quality gate is soundness theater", "S1 never stalls waiting / S2 one-use advisory", "Local controller ≠ githubnext/localjev", "purple telemetry = consumed not arrived", "seed = geometry not async replay", "20% starting gate still soft", "no pixels to either provider", "OCR+AX observe-score-act / typesafe-computer-use", "never send screenshot to frontier for the decision", "overlapping CU options = false low confidence", "split kind/item/site", "155× one-screenshot ≠ Harbor taskset", "decision ≠ answer-reader capture", "ASR observe-score-act / jev-voice-browser", "partial-speech VOI / free-text waits", "spoken confirm ≠ hard auth", "numbered overlay without another model", "wrap-as-execution / AgentGhost ALLOW ASK DENY", "rules first then Jev remainder / ASK throws / fail-closed", "reddpy/AgentGhost ≠ jwen5419807/agentghost ≠ vventirozos", "JP genre atlas / studio_yebisu / stars ephemeral ≠ eval", "Jev Clearly Explained / akshay_pachaar / LLM hammer", "schema-safe ≠ correct / 200× 400× TypeSafe ceiling", "questions-as-code / shadow first / not a TypeSafe how-to", "proposition ≠ embedding / contrast-set", "boolean composition of soft Nouls / AND OR NOT", "uehaj/jev-semgrep ≠ semgrep.dev", "meaning-grep dedicated fold / not a gate", "decision-validated UI / Jev never authors text", "decision-as-assert / jevtest ambiguous band", "typed decisions drive UI / jev2ui", "hybrid S1 closed verb menu / anima3", "pointer-not-generator search / JevFind", "jev-frontier-bench ≠ frontier-100", "product bakeoff ≠ architecture duel / GLiClass", "four engines same questions / majority floor", "authorship named escape / not evidence", "ha-switchboard HA remains execution", "n8n classify/route/score / Low Confidence", "fast-jev-compaction-pi ≠ pi-jev-compact ≠ pi-jev-compaction", "jevloop full-distribution optimizer / no LLM in the loop", "laya-vision SmolVLM / score untrained", "Cerebellum-2B /v1/decide ≠ TypeSafe / wire-compat vs agent-routing", "laya-grounded not drop-in / Platt not temperature", "GestaltLabs/Jeff-1 ≠ logan-markewich/jeff / acc vs ECE n=9730", "stanley-code empty findings ≠ approval / human promote", "findme ≠ JevFind / NL memory beam-search FS", "jevsubrouter price workers not conversation / counts ≠ dollars", "feelings .feels() default 0.5 is Noul-0.5-never-rounded / ≠ hunch ≠ Probably", "apa-agent-harness ≠ AntonioCoppe/jev-harness / unpublished npm", "grok-bot-jev skill cannot force a bot that ignores it / A/B proxies not tokens", "Essentiel-Jev never authority / human every action", "enzo-mcp independently falsifiable claims / ≠ jev-sift", "pigeonhole OTHER skip / decision-as-filing", "jev-reliability Nothing about accuracy", "clduab11/jev-test ≠ realZachi/jevtest / Nothing runs yet", "jev-rag-benchmark Jev wins is not an assumption", "dairui1/jev-lab ≠ BrendanH18/jev-lab", "jevmail gmail.readonly / mailjay archive/trash", "ZHUBoer/ego-jev reserved __none__", "runWorkflow completed ≠ success", "jsort scores are relative", "Noul not Choice for scale", "groundedness-judge-bench native vs schema-guided", "implicit_true included in yes", "jev_playground 0 promotions", "routing-backtest 0.0447%", "yuyang2230/jev-agent-skill jev-1.13-free", "jev-techstack-classifier stack_config.json", "s1_ruby collapse late", "undecided? abstain", "2389-research/judgement license null", "confidence ≠ winner p", "typesafeai-sdk-community not a new species", "tpellet/hunch exit 3", "never-execute list", "jev-file-search scores not calibrated accuracy", "jev-linkmap Jev never sees S2 prose", "muhammedilyasy/jev-mail metadata only", "tidy none-of-folders stay", "tab-bouncer pinned/audio/current never closed", "lkclean Show fail-open", "jev-yt-time-saver Show anyway", "ORIGIN pause-if-no-Jev", "validResponse sums-to-1", "jev-crawlers risk bands never raw boolean", "jevbrain AUTO_ACT is not a Noul", "judgekit YAML classify/score/route/verify", "typed-judge-kit verdict-in-code", "alsoleg89/decide packing VOI", "0.8 ≠ 80% accuracy", "Jev-Calibration Platt ECE 0.117→0.052", "jev-calibration-arena never acts", "ctmx/openrouter-jev-mcp Decision-as-Plugin", "FrancoisChastel/jev-code ≠ npm jev-code", "claudecode-jev-marketplace fail-open not hot path", "pedroknigge/mcp_jev packs not ask_jev", "cyrusasco/typesafe-mcp noul deadband 0.35–0.65", "codaaiteam/jev-skill jevtypesafeai.com ≠ TypeSafe", "hermes-switchyard ≠ hermes-jev-router ≠ hermes-plugin-jev", "nanoprune 2.8MB ECE 2.58%", "smartdio/jev-browser-agent ≠ ZHUBoer/ego-jev", "Dakai/omp-jev-web DONE ≠ proof", "hari007sh/jev ≠ dannote/jev", "0thernet/system-one-skills deterministic verify", "typed-gate band [0.40,0.60] is refusal", "pi-jev-gate fail-closed; choice is the verdict", "Foq ~25ms/2.2GB local", "rev prefill-only + HF jev-0.5b", "robfrase/jev planning memo", "typesafe_agent_gates 27/27 / 31/31", "EpicEric/safe-sh static remainder", "pastepilot Confirm before act", "Jev-Reranker live Jev not yet measured", "sessionwise opt-in relevance", "jev-search pointer sieve", "400ms Salesforce WebMCP", "typesafe-scheduler-diagnostics advisory", "droidjev screenshot-free", "Tewoto1 jevcu planner still writes", "ha-conversation-jev Jev→Grok", "dsh-jev can only gate", "jev-classification-benchmark specified not run", "jev-luna-pagerduty p≥0.50", "meldltd/meldecision laya-go ONNX", "laya-doom never pixels", "logixism/laya-api empty README", "akpsahan/laya ≠ Archer", "choxos/jevchess engine owns truth", "jev-drive sim not AV", "story-arc Jev never authors", "jev-hs-assistant HS6", "golergka/jev-plays-starcraft-2 UI-verified ≠ API Victory", "awesome-jev-use-cases catalog", "Nibir1/typesafe-go ≠ official", "fingerprint after redact", "recall vs decide", "publish fingerprints+answers", "CI replay as Harbor cousin", "Cache hit ≠ correctness", "hyperspaceai/jevcache ≠ kushals256/jevcache", "human labels only", "score never auto-accepts", "production capture flywheel", "sutro-sh/jev-align ≠ caiovicentino/jev-align", "guidance ≠ hook", "catalysts ≠ summaries", "compile-time System One", "unofficial ≠ TypeSafe", "format_version modernbert-jev/1", "Argos1111/jev_local ≠ us/jev-local ≠ kunchenguid/local-jev", "LFM default ≠ ModernBERT backend", "Nemotron ≠ TypeSafe Jev", "not a calibrated replacement", "djev-dev complements djev-spark", "images as Choice options", "Laya essay numbers *theirs*", "Router/OOD confidence", "hosted bootstrap ≠ silent TypeSafe", "difficulty + policy thresholds + JSONL trace", "jev-codex-pilot model + reasoning depth", "keep/shadow/hybrid/reject", "quarry evidence projection", "Frank-ZY-Dou/awesome-jev robotics/3D/control", "one-dollar-tahoe TypeSafe Jev defense eval", "jevguard calibrator/cache/escape", "jev-ci-selector CI shadow mode", "llama-jev llama.cpp replica", "petercr/jev-orchestrator ≠ FleeexCorp/jev-orchestrator", "seb4ez/jevguard ≠ AseemPrasad/JevGuard ≠ pablozr/JevGuard", "webNeat/llama-jev ≠ WiktorB2004/llama-index-jev", "OpenCode jev-pruner context sieve", "observe→score-candidates→prune", "jev-zen / jev-1.13-free", "zen-chat ≠ Noul", "fail-open original", "keepScore >0.1 floor", "host port of tamaratran/jev-pruner", "indiejoseph/opencode-jev-pruner ≠ nrdz-labs/fast-jev-opencode", "jev-webagent-bench empty stub", "Kiln-AI/jev_jsonschema noul_threshold 0.5", "NSStudent/JevSwiftSDK unofficial", "GLiNER2 native Apple path", "unofficial Swift/Core ML GLiNER 2.5-small", "entity spans + confidence", "not Choice/Score/Noul", "not TypeSafe", "label descriptions as schema", "on-device ANE economics", "honesty locks", "shershah1024/gliner-native-runtime ≠ Fastino", "≠ gliner25-compaction ≠ gliner2-ultrafast ≠ Eran-BA/Jev_from_GLiNER2 ≠ NSStudent/JevSwiftSDK ≠ jevmlx", "default threshold 0.1 still soft", "soft Noul ≠ hard safety", "Decision Graph Protocol frame→assess→commit", "app retains permissions/effects", "Jev-first assessor-neutral", "guarded commit / receipt/next frame", "assessment batching", "hard-gating DGP as safety theater", "numerous-com/dgp ≠ TypeSafe official", "jegrep calibrated path+range Nouls", "no embeddings/index/daemon", "~$0.01–0.03 typical", "agent --json", "can1357/jegrep ≠ Bentlybro/jevgrep ≠ uehaj/jev-semgrep", "Archer-arch fidelity", "kev family OOD 0.76–0.77 vs Jev 0.86", "block-causal isolation", "pointer/readout CE-trained", "/v1/systemone drop-in", "replica honesty", "cost-sensitive decision theory × System One probabilities → control flow", "thresholds derived from costs not hard-coded", "YES / NO / UNSURE from cost_false_yes / cost_false_no / cost_human", "auto-batching same-object questions", "Kungie/gut ≠ tpellet/hunch ≠ carldaws/hunch", "judgment vs generation", "deterministic execution after probabilistic judgment", "exactly one app-owned callback", "explicit uncertain branch", "Illusion47586/judge ≠ lexingtonhibiki/judgekit ≠ Ascurse/typed-judge-kit", "variable-N option scoring as the trainable object", "dynamic candidate bags not fixed label sets", "zwliJay/jev-forge ≠ NanoJev", "open replica economics / latency vs closed Jev", "NAR local drop-in", "wfzyx/von late-catch HIGH", "competing NAR claims / replica honesty", "typed judgments vs chat judges on guardrailing", "ishaannk/llm-vs-jev cross-note only", "deeper integrity fold is rh-guard", "nothing wins outright", "can be argued out of guarding"", "Jev IS the if-statement", "judgments/probabilities drive branches", "text model only writes prose", "interpreter owns variables/loops/budgets/replay", "otherwise maybe / confidence gate", "chaos samples after the gate", "southpolesteve/probably ≠ carldaws/hunch ≠ feelings ≠ Kungie/gut ≠ Illusion47586/judge ≠ tidymodels/probably", "133★ / forks 10 live", "build calibrated classifiers from human feedback", "retrieve by relevance not resemblance", "one calibrated yes/no per memory in one request", "pointer mode 17/18 19/20 *theirs*", "embedding resemblance misses the allergy", "samdotmak/jev-recall ≠ jev-search ≠ jev-sift ≠ carryforward ≠ chopratejas/invalidate", "memory leases ended by new evidence", "six Nouls then fixed rules in code", "0 of 157 false invalidations", "questions/plans/directives are not evidence", "unsure → review queue", "host keeps the store", "name↔body / comment truth / test-claims", "mizchi/jev-lint is mizchi/jevlint rename", "no shipped rule has severity error", "~1 in 5 findings wrong *theirs*", "mizchi/jev-lint ≠ huntedman/JevLint ≠ MichitoSugawara/jev-lint", "JSON Schema → typed JSON via Jev", "noul_threshold 0.5 decoder not a proof", "IncompatibleSchemaError lists every bad property", "on-device Laya CoreML ANE", "~5 ms P50 short decisions", "189/189 FP16 checkpoint parity", "10× not achieved", "mizorewww/laya-coreml ≠ gliner-native-runtime ≠ jevmlx ≠ NandhaKishorM/laya", "softmax over allowed tokens ≠ Noul", "question-first cache", "Micha0827/snapjudge ≠ githubnext/localjev ≠ jevmlx ≠ cendress/SnapJudge", "Jev-first Pi agent loop", "slow-LLM fallback", "explicit action menu / CandidateSource unimplemented", "62 tests wiring not quality", "direwolfiy/JevPi ≠ standardagents/jevpilot ≠ pi-jev-control", "resume-screening bias audit methodology", "name×resume factorial independent Nouls", "callback determined by resume quality", "mean-probability name gaps operationally negligible", "natemoo-re/bias-bench ≠ BBQ", "Plan/PRD panel → code-owned pass|review|block", "cheerleading out of scope", "austindixson/planalyzer ≠ single-goodness Noul", "cost-aware multi-model routing/escalation", "decide vs do", "successful-task cost", "cannacre8ive/switchboard-ai ≠ ha-switchboard ≠ hermes-switchyard", "frozen-protocol zero-shot bench", "TypeSafe Jev vs PrismNLI vs Laya", "contamination caveat", "elcronos/jev-vs-open-decision-models ≠ JevBench ≠ DMB", "context-window admission control", "VOI gate which tokens are worth the expensive model", "fail polarity per lens", "on small inputs lenses lose money", "cvsgireesh/jevusher ≠ jev-sift ≠ winnow", "typed decision control plane", "receipt ≠ authorization", "historical-v0 zero retained cases", "MokiMeow/jev-fabric ≠ jev-forge ≠ dgp", "live 15-dim typed rubric re-score per pause", "scoring economics exemplar", "OpenJev/Codiv ≠ TypeSafe hosted", "jose-troche/live-rubric ~$0.000004 desc / ~$0.000006 README", "adversarial pre-registered Jev eval", "28 predictions before data", "123,805 requests", "confidence does not track ignorance", "polite injection 65% / crude 0%", "willkelly/jev-evaluation ≠ jevals ≠ jev-baselines-eval", "provider-neutral Elixir/BEAM Noul/Choice/Score SDK", "class infrastructure", "nshkrdotcom/system_one_sdk ≠ typesafe_sdk ≠ dannote/jev", "question-linting of Jev questions themselves", "nine jaggedness rules, no API key, no labelled data", "static lint ≠ measured separation", "yodablocks/jevq ≠ tenbin ≠ JevLint ≠ commitjev", "open-weights Laya as class exemplar (binding)", "Nx/Bumblebee runtime", "host chooses backend", "ChristianAlexander/laya_ex ≠ system_one_sdk ≠ dannote/jev ≠ NandhaKishorM/laya", "on-chain/edge Laya deploy", "parity_verified stays false", "model output never grants Tx", "humandebri/IC-Laya ≠ laya_ex", "auditable weekend replica", "Jev outputs never used for training", "soft human-vote distributions", "unpaired 0.577 vs 0.727", "agilabs-ai/jev48 ≠ JevBench ≠ Mapika/decider", "adversarial dual-judge / framing attack surface", "comparative framing is the usable judgment", "prior injection crowds out evidence", "copyleftdev/ember ≠ ember.js", "Laya specialist fine-tune pipeline", "training still GPU-pending", "PIXELZX0/XERON ≠ convaiinnovations/laya", "Hub Laya replica drop", "daliborsb/laya ≠ convaiinnovations/laya ≠ NandhaKishorM/laya", "System One student distillation corpus", "gold is programmatic", "teacher is closed-API clone", "do not distill Jev as teacher of record", "MagaBitmex/jev-4b-distill-data ≠ missing student checkpoint", "non-LLM VIN System One", "planning depth not chat", "lewislululu/jevon ≠ douglance/jevon", "source-bound evidence checks", "local quote mismatch needs no API", "exit 0 ≠ claim truth", "WaynezProg/jev-kit ≠ jonathanavis96/jev-kit (Airlock) ≠ jev-use ≠ jev-mcp", "independent System One evidence catalog", "scores not one leaderboard", "no external record currently reproduced", "TokenTrim no-Jev matched hybrid 62.4%", "reachjalil/system-one-bench ≠ mallahyari/system-one-benchmark", "21 tasks · 134 items · 208 questions", "scenes from public GitHub contracts, not production logs", "SivletLabs/jev-eval ≠ willkelly/jev-evaluation ≠ 4esv/jev-eval ≠ xxkuboxx/jev-eval ≠ onlyoneaman/jev-eval ≠ dayhaysoos/jevals", "option isolation (sibling-blind)", "permutation-equivariant", "Hub OWNER not published", "nafisazizir/hev ≠ jaredpalmer/kev", "frozen local LLM logits, no trained decision head", "residual-head 9,222-param decreased 73/96→67/96", "confidence = 1−normalized entropy, not P(correct)", "yuki-oshio/mini-jev ≠ r-ms/mini-jev", "Jev classifier as autoregressive next-token predictor", "ChatJev-style soundness theater", "erik-dunteman/ChatJev ≠ dannote/jev ≠ jev-gpt", "calibrated decision head × AlphaProof value head", "implementation-layer isomorphism, semantic difference", "timeout = censoring", "do not launder Noul as proof", "parallel rank-prediction vs serial selection", "independent questions can conflict", "zzzzzec/jevsort ≠ keltokhy/jsort", "curated open System One ecosystem catalog", "rupeshpoojary9/awesome-open-system-one ≠ AnotiaWang/awesome-jev", "arXiv paper radar with Jev relevance scoring", "ranking ≠ calibration / 0.5 still soft", "fail-open failed evals not marked seen", "train calibrated ~27M from scratch", "typed Q→prob dist / one forward pass / no LLM decode", "hyusi2003/MiniSystemOne ≠ Colvin0315/MiniSystemOne", "description-only stub / size 5", "ESCI hard probe fails four of six", "jev_bool ECE 0.242 inversion 0.255", "do not re-fold §60 six-gates as new", "jobbyjev one-request-per-company from batch-size result", "find/design/evaluate TypeSafe Jev decision loops", "karanb192/jev-architect ≠ samtay32/jev-system-architect", "Jairik/jev-distiller size 1", "distill-Jev UI stub / do not distill Jev as teacher of record", "post-launch scored use-case map / Jev self-scores then human curation", "licensedsaucer9-web/jev-opportunities", "Jev-inize a use case into classifier/router", "gavinHuang/jevinize → simple-jev not TypeSafe", "featherless-ai/simple-jev", "compare saved decisions / same label can still change the branch", "VihaanAgarwal/jev-diff ≠ Saik0s/diffusiongemma-jev-macos", "not tested with a live Jev API key", "constrained logprob + temp/Platt ≠ Noul", "OpenJevPro pastes openjev-sglang JevBench as own", "zhangcy122/OpenJevPro ≠ IamBusy/OpenJev ≠ ekzhang/openjev-sglang", "PolyForm Noncommercial", "SmolLM-135M / sub-70ms / 0 output tokens", "demo P(True) 0.5052 / Choice conf 0.2872 / Score conf 0.0055", "README claims MIT / GitHub license null / no LICENSE file", "patelvishwa112/jev-system-one-rlcd ≠ arnabgho/rlcd-lite ≠ blackwood-rlcd", "source-backed Awesome Jev radar / 306+ commit-pinned", "logicrw/awesome-jev-projects ≠ AnotiaWang/awesome-jev ≠ yibie/awesome-jev ≠ cobanov/awesome-jev ≠ rupeshpoojary9/awesome-open-system-one", "auto GitHub sync / Issue-only submissions", "hashed n-gram encoder / rival-aware attention", "olanotolu/jevbetter vs jevlike starter", "synthetic hard menus top-1 0.916 vs 0.873 / ECE 0.0182 vs 0.0367 / 40 vs 4608 menus/sec", "shuffled-context control 0.335", "Turn any open LLM into System-One Jev", "uspraveen/Jevify ≠ Mintzs/jevify ≠ gulagala001/jevify", "Jevify-any-LLM architecture probe", "description-only stub / size 0", "Train encoder-only calibrated decision models from a task sentence", "Exu is a toolkit, not a method", "strictly proper scoring rule", "Pre-alpha", "Ruivalim/exu-base", "scratch-trained calibrated decision model", "typed Q → probability dists", "Colvin0315/MiniSystemOne ≠ hyusi2003/MiniSystemOne", "no published weights download URL", "90.5 seconds / 29.2% pipeline evidence", "p_i/p_j independent of other candidates", "Recipe for calibrated decision models — small model out", "init → synth → train → eval → serve", "91.1 % / ECE 0.022 *theirs*", "Jev zero-shot 75.1", "scienthoon/luce", "Put Jev's three headline claims on trial", "0.5B local GPU", "46x speedup / accuracy identical", "ECE 0.624 sentiment catastrophe", "bigger model worse calibration", "RichardoMrMu/jev-mini ≠ yuki-oshio/mini-jev ≠ r-ms/mini-jev", "System-1 decision engine for local LLMs", "structured choices only", "JSON parse of generated text ≠ Noul", "TypefAI JEV / Journal Entry Voucher", "tapsin/jev-local ≠ us/jev-local ≠ Argos1111/jev_local", "Jev 1.13 reward-model eval across 8 benchmark tracks", "40,940 examples / 0 API errors", "RewardBench v1 92.58%", "Precise IF 50.63%", "goya4140/jev-reward-model-evaluation", "Scaffolding in progress", "Jev vs LLM support-ticket routing", "static + live decision bench", "TypeSafe's own published benchmark", "illustrative simulations, not live API calls", "JevBench v1 — smart/cheap/fast/reliable", "I/C/S/K 25% geometric mean", "classifier.dev fast tier 84.8 is Jev behind its own API", "do not re-fold §78 v1.2 board as new", "Laya (421M) 70.1 now on board", "Zero-shot/few-shot LLM routing", "hard budget filter before Jev", "Jev never asked to perform budget arithmetic", "Jev judges the next state, XState enforces transitions", "simulation uses synthetic keyword fixtures", "catalog gravity", "v-modal/awesome-jev-tools", "★339 live REST", "curation is not endorsement", "crawler-maintained directory", "Daily GitHub + npm sweep, human-merged", "RadRebelSam/awesome-jev ≠ AnotiaWang ≠ yibie ≠ cobanov ≠ logicrw ≠ v-modal", "HF peft SPLADE/BGE reranker", "rdxtremity/jev-reranking ≠ carlaiau/jev-reranking", "query-side encoders, not a Jev replica", "ONNX System One Qwen3.5-4B scorer", "source:pngwn/system-one-qwen3.5-4b-scorer", "CC-BY-NC-4.0", "temperature 1.75", "transformers.js AutoModel cannot load this graph", "Consistency benchmark Space", "This Space contains no benchmark result yet", "12-case plumbing fixture", "Benchmark-driven Jev router and judge", "cheap alone is not success", "Jev does not write, sum prices, or claim accuracy %", "Sol 94.2 / Luna 83.9 / Jev path 89.7", "19.2% Sol / 62.3% cost save / 4.5pp miss of 2pp non-inferiority", "p50 latency worse than Sol due to routing overhead", "erendikmenn/jev-llm-router-benchmark ≠ jev-rag-benchmark ≠ ryantsai/jev-llm-router", "Express + node:sqlite", "mock and Jev decision engines", "previous_ticket_count >= 3 is code", "MIN_CONFIDENCE 0.6 still soft", "substring false positives", "aesaganda/jev-ticket-router ≠ SarathChandraBellam/jev-vs-llm-ticket-router", "Universal Figure & Diagram Router", "confidence ≥ 0.85 hard-gate is theater", "generative AI banned from scientific plots", "six visual branches", "hoangngochuong24947-gif/jev-figure-router", "human-labeled (state, question, label)", "166,054 rows / 22 configs", "soft_label for human uncertainty", "Praveenrajus/jev-bench ≠ fstandhartinger/jevbench", "ternary bonsai System One GGUF", "openjev's mechanism, Bonsai's weights", "Hub does not ship weights", "100/100 easy T/F is not Harbor", "label_mass ≠ correctness", "stock llama.cpp Q2_0 silently gibberish", "NicolaiMTLassen/open-bonzi-jev ≠ NicolaiLassen", "transformers.js DeBERTa ONNX", "source:com-kotobalabs/open-jev-deberta-v3-large", "temperature 1.05", "AutoModel from_pretrained works", "onnx-community/open-jev-deberta-v3-large-ONNX ≠ system-one-qwen3.5-4b-scorer-ONNX", "107★ densify", "GH 151M vs README 149.6M", "PR #1 now closed unmerged", "do not re-fold §71 claim-audit as a beat", "typed decisions, RLCD, confidence-gated routing", "structured ≠ correct", "mock not live API", "26 tests", "wjdjdakf17/jev-study ≠ baekenough/jev-study", "bonzi-27b-v2 / ternary-8b / 27b-v1 GGUF family densify", "WANLI-256 74.6% / 65.2% / 71.1% *theirs*", "Bonsai 1 27B Q1_0 runs on stock llama.cpp", "ternary still needs PrismML fork", "hf:heman10x/openJev-verdict-2.0 twin tokenizer-only", "OpenJev Vision image classification + uncertainty", "CLEVR-4 held-out joint 0%", "hfdataset:IamBusy/OpenJev-Vision-Research-v0.1 12,832", "294,912 derived targets not independent samples", "Laya multilingual ONNX WebGPU typed-decisions port", "63/63 selected answers / 5.1e-4 CPU / 1.2e-2 WebGPU", "UpHash-Network/mini-jev is yuki-oshio transfer", "jev-injection-bench 11,900 labelled prompts", "Jev best ranking / Haiku better ECE 0.021 vs 0.058", "0.5–0.9 band is where Jev's numbers do not mean what they say", "Prompt wording moves panic 28%", "manojlds/jev-dspy-bench ≠ dspachos/jev-dspy ≠ jmanhype/jev-dspy-lab", "Jev agreement is similarity, never ground truth", "no aggregate quality grade or merge gate", "AbstentionBench-on-Jev rank 1 of 20 vs 2025 field", "question-asymmetry", "forward-looking 0.465 never extreme", "openkev calibration layer not a runtime", "ECE vs coverage independent", "select_threshold returns inf", "escalation catches uncertainty not ignorance", "misakaikato/openkev ≠ jaredpalmer/kev", "pdf-race Docling→Jev vs Gemini", "parser owns the wall clock", "12/12 tie is a tie", "titles selected not generated", "flopcheck 16 calibrated tweet judgments", "mechanical tells in code", "ZeroX-01/jev-atlas ≠ Zaious/jev-capability-atlas ≠ gorock007/jev-atlas", "Laya calibration lab Gradio MCP", "T never changes argmax", "confidence ≠ top-label p", "easy probe set refused", "40–48 rows too small to ship T", "Gemma-4 26B-A4B jevify classification+calibration", "LoRA adapter twin not independent eval", "Gemma-4 E4B jevify", "E4B LoRA stub card", "kushalpatil/jevify-gemma4 ≠ Mintzs/jevify ≠ gulagala001/jevify ≠ uspraveen/Jevify", "GH kushalpatil07/jevify 404", "PAWS 0.580/ece 0.288 is the weak cell", "smaller E4B slightly better OOD ECE than 26B-A4B", "Hub jevify merged LoRA ships weights", "bonzi Bonsai-8B v1 GGUF densify", "Bonsai-1.7B v1", "Bonsai-4B v1", "WANLI-256 64.5% / 60.2% / 52.0% *theirs*", "rank #4 / #5 / #6 of 6", "JulesHuisman/jev-eval scaffolding / README SHA c356a584 (was empty e69de29b)", "JulesHuisman/jev-eval ≠ SivletLabs/jev-eval ≠ willkelly/jev-evaluation ≠ 4esv ≠ xxkuboxx ≠ onlyoneaman ≠ dayhaysoos/jevals", "7 bands 6/10 vs 40 bands 0/10", "source receipts + confidence slider re-policy without re-inference", "32/32 synthetic is smoke not production", "classify HF datasets across typed semantic dimensions", "roadus2 watch misspelling; lock roadius2/ultra_laya", "ultra_laya REVIEW defects", "default branch claude/laya-jev-review-gg5ppo", "XNLI EN 88.3% ECE 0.032 → RU 77.3% ECE 0.096", "Δ −11.0 pp [−14.2,−7.8]; ECE +0.063", "MASSIVE no detectable difference at n=600", "confidence is function of p_max (r=1.000)", "pointer-not-generator 400 human-authored responses", "proposed ≠ authorized", "FewRel 160: Jev 85.0% vs lexical 13.125%", "gated 100% (95/95) coverage 59.375%", "J++ composable semantic computation language", "judge-jev 0.5 still soft", "947 repos scored; A 273 / B 302 / C 372", "LLM rubric ≠ benches", "No benchmark winner is claimed", "phishing: naive 62.6% vs regex 91.8%; 5-atomic + LR 95.0% *theirs*", "AITuber tension ±15", "README npm global; repo is Rust", "git-confess code owns counting/blame/ratio", "httpx exhibit 11% (13/119) *theirs*", "90d trend +12.40% vs random +12.75% vs BH +41.71%", "5m win rate 25%", "Awesomejev 656 entries / 38,160 stars", "tracker likes 64 (+4) lastModified UNCHANGED", "Laya present; Blackwood ABSENT; Archer still promised_not_landed", "Blackwood tracker ABSENT; likes 2 gated manual", "r = c - p_a", "ECE 0.021; acc 0.807 vs warmup 0.746", "Independent primitive", "11.57s vs 54.10s · 4.67× · 120/128 *theirs*", "default path is pretrained Gemma probs not trained RLCD head", "GH Meanblock 404; lock leesk212/JEV-CPU", "softmax over letter slots ≠ Noul", "WANLI 0.741 vs openjev v2 0.77 *theirs*", "3-way NLI ≠ Noul", "priority 0.464 = majority floor", "banking77 contaminated", "raw margins not probabilities", "do not distill Jev as teacher of record (they distilled Haiku)", "“0.9 is not one number”", "ranking ≠ calibration", "banking77 0.8–0.9 stated 0.86 actual 0.73 over-confident *theirs*", "≠ Praveenrajus/jev-bench ≠ fstandhartinger/jevbench", "$0.0000153–$0.0000226 vs circulating $0.0004 (~20×)", "Score is 0..n-1 expectation not 0–1", "Noul has no confidence field", "TCP floor 198.8 ms", "type reliability is not a reason to choose Jev (json_schema 5/5)", "gateway tax not one number", "Function-only 5/8 vs hybrid 8/8", "4/8 without Jev", "8 designed cases not conversion lift", "200-row pilot Jev 86.5% 173/200 vs Gemini Flash-Lite 86.0% 172/200 vs Pro 87.0% 174/200 *theirs*", "not a ranking", "情緒測謊器", "8-example Jev vs GPT-5.6 Sol ~64× cost 5.4× latency *theirs*", "synthetic; no inference", "≠ JevBench v1.2 §78", "Judged 3317 / listed 2560", "Jev judges, code applies policy", "APA “microsecond policy / zero hallucination” overclaim", "Client-side quiz; pointer from held docs; scanned-PDF warn", "Jev judges / agent reasons / user decides", "selecting an option is not permission to implement", "pattern exact, judgement must clear floor", "no matching pattern → no model call", "not a correctness oracle", "Spec vs artifact remainder", "treating 0.85 as 85% / minProbability hard-gate as Harbor", "VERIFY acquires discriminating evidence, never same-pool confidence-only rescoring", "fast/full/max are ceilings not sizes", "Solar writes, Jev chooses NEXT ACTION", "do not reopen or amend PR #23 or #24 or #25 or #26 or #27", , "Calibration is not alpha", "NO CURRENT ALPHA CANDIDATE", "ΔR² approximately +0.00084", "Brier 0.2131387", "ECE 0.0421875", "Adding Jev probability to deterministic volatility improved Brier by only 1.4058e-05", "default 0.5 keeps zero non pinned", "keepResult median 0.14 to 0.17", "keepCall median 0.28 to 0.35", "usable range is about 0.10 to 0.25", "7.8% to 57.9%", "judges results it never sees", "task-finish eval not built yet", "$0.002 per compaction", "slavadubrov/sgr-judge-bench ≠ slavadubrov/jev-judge-bench", "Jev 108/120 $0.083 0.34 s", "Luna SGR 114/120", "paired Jev accuracy-difference intervals include zero", "not evidence of equivalence", "GLM SGR 26/120 93 format failures", "Terra-planned Jev hybrid 55/120", "rule-based by default, optionally Jev-backed", "empty README", "missing key cannot break the experience", "prefill plus exactly one decode", "softmax over A/B/C ≠ Noul", "BBQ 9,053/10,000 (90.53%)", "ECE 0.0890", "Mean confidence 0.9943", "overconfident", "score and noul not implemented", "DGUI 12 rows (was 6)", "INSTRUCT 119 rows likes 2", "encode the state once, decide everything in parallel", "0.740 accuracy against a 0.508 majority", "ECE 0.047", "fine-tune's advantage ends where its 384-token training data does", "jasonkneen/open-jev ≠ pngwn/open-jev", "same sha d41dc3cd", "Space does not call Jev", "recomputes routing from saved probabilities", "200-case Jev 97.0% / 100.0% / 95.0% / MAE 9.22", "synthetic repository benchmark", "Jev evaluations are advisory", "YehuiTang0316/jev-nlgrep ≠ Bentlybro/jevgrep ≠ can1357/jegrep ≠ uehaj/jev-semgrep", "default threshold 0.8 still soft", "40-line windows cannot prove whole function", "token-native sequential start/end Choice", "Gemini/Haiku stubs not configured yet", "handful of hand-written examples, not a benchmark", "Jev judged exactly what it was given", "laguagu/jev-skills ≠ laguagu/jev-evidence-lab ≠ Pleo2/awesome-jev-agent-skills", "contract_passed is not a claim of guaranteed factual truth", "Wilson lower bound 0.85 floor", "fixture mode no savings claim", "SemIf 2207★ (+21 vs §110 2186)", "jevlike 1043★ (+5 vs 1038)", "TypeAR 15★ (+1 vs 14)", "AnotiaWang 97★ (+1 vs 96)", "yibie/awesome-jev 506★ (+16 vs 490)", "Laya likes 822 (was 802)", "tracker likes 64 flat, lastModified UNCHANGED", "do not reopen or amend PR #23/#24/#25/#26/#27/#28", "Heman10x-NGU/Verdict-open-jev ≠ Heman10x-NGU/openJev-verdict-2.0", "TF-IDF + LogReg ECE 0.0207 vs Jev 0.1440", "Verdict-open-jev 48.07% vs Jev 90.80%", "abstention combined recall 10.00%", "p50 35.58 ms", "K=25 (maximum capacity) 72.00%", "0.85 coverage 84.60% selective risk 1.18%", "26.1× faster than standard Qwen JSON generation", "Jevify 90.0% / 167 ms CUDA graphs disabled", "Finding 1: Brier on stated confidence alone is a trap", "grpo_rlcr 0.78 / ECE 0.084", "reliability 0.007 but resolution 0.000", "27 900 schema-driven decisions", "13 600 / 13 600 questions", "candidate mass min 0.99999624", "22 configs · 166,054 rows · 4 calibration-gold", "sha a39eba3f", "Student B MAE 0.148 / Pearson 0.836 / 86.0%", "pngwn/open-jev-laya-bench README 404", "sha 9f69c742 likes 2", "HDFS 0.9933 (745/750) / retain 0.0084", "BGL ERROR/FATAL protection 1.0000", "2,479 / 2,500 HDFS uncertain", "cache hit 0.9648 (2412/2500)", "$0.153936 estimated", "E2 recomputes from saved probabilities", "Space sha eda59e0a", "MASSIVE English 0.783 / Khmer 0.033 / Hindi 0.133", "40–48 rows too small to ship T", "T never changes argmax", "siren2345/jev-single-decode-transformers ≠ siren2345/jev-single-decode", "Split Transformers experiment from llama.cpp runtime", "tanayvasishtha/jev-lab ≠ dairui1/jev-lab ≠ BrendanH18/jev-lab ≠ yibie/laya-jev-lab", "Four experiments stress-testing TypeSafe's Jev: calibration, bundle bias, label bias, and ensembling", "second pass must be $0.00 from cache", "The pages never call Jev", "Gemma 4 31B 77.0% / Jev 1.13.0 61.4% / Laya 322M 0.0%", "restriction state 95.0% against 84.4%", "None of the systems are particularly good at knowing when to stop and ask", "They skip the question and call a tool directly", "100% schema pass", "six-field joint 48.8% vs 72.8%", "ywchiu/jev_benchmark ≠ Running-Dolphins/jev-bench ≠ Praveenrajus/jev-bench", "ACT / REVIEW / FALLBACK", "A provider failure, timeout, malformed output, or missing answer is **not** a policy outcome", "confidence is descriptive provider output, not a substitute for probability", "Quality denominators include only valid scored answers", "an exact halfway tie chooses the lower level", "aiwithenoch/Jev-Skill ≠ simplosophy/jev-skill ≠ laguagu/jev-skills", "The local path does not claim to turn a smaller checkpoint into Jev", "Low support becomes decision: \"review\"", "MIT-0 SPDX NOASSERTION", "current-llm", "结构兼容，不是 Jev 模型能力", "altryne/jevify ≠ Mintzs/jevify ≠ gulagala001/jevify ≠ uspraveen/Jevify", "Find where Jev belongs. Design the questions. Measure the difference", "TypeAR-AI/TypeAR 301 → TypeLLM/TypeLLM", "TypeLLM/TypeLLM 16★", "SemIf 2241★ (+34 vs §111 2207)", "jevlike 1051★ (+8 vs 1043)", "AnotiaWang 98★ (+1 vs 97)", "yibie/awesome-jev 525★ (+19 vs 506)", "Laya likes 864 (was 822)", "tracker likes 67 (+3 vs 64)", "lastModified UNCHANGED `2026-09-20T04:29:16.000Z`", "do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#32", "hysteresis enter/exit / replay policy without inference", "calibration does not compose / hop-ECE permutation-invariant", "equal-width vs quantile ECE / ranking ≠ calibration", "Qwen2.5 ≠ Archer / Qwen 3.8 sparring ≠ Archer / Qwen/Qwen3.8-27B ≠ Archer", "Deferred Crispification / TCE / AMS", "g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev", "pd.cut equal-width vs jeval quantile", "A hunch is a probability with a policy attached", "soundness theater / measurement theater / hourly 0843", , "Jev Capability Resolver / NiazMorshed2007/jcr", "one tool nested capability tree / returns context / does not execute", "skills vs capabilities / workflow+judgment vs operations", "format independent of Jev / proposed open standard", "JCR_BAND_RATIO 0.6 is application policy / soft scores ≠ hard gates", "routing ≠ permission / docs ≠ authority to run", "sol-vs-opus5-20 lookup+explain / n=1 / Not Harbor task-execution", "wall-time mixed / Sol slower with JCR in 19/20", "NiazMorshed2007/jcr ≠ skill-broker ≠ skillranker ≠ jev-sift ≠ jev-lens ≠ jevusher ≠ jev_select_capability", "do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34", "notes.md §116", "copy the SemIf/MLX installer?", "quote 5.21× as beating Jev?", "treat 0.845 as a TypeSafe replica?", "collapse SemIf into kw2828/zhihz/semif-rs/semif-serve", "softmax over options as a Noul", "llm prompt to jev primitives", "conversion assistant not equivalent behavior", "heuristic conversion ≠ calibrated Noul", "alexwestco/llm-to-jev ≠ altryne/jevify", "user-provided 0940 / notes.md §118", "judge ≠ actuator", "candidate_mass", "softmax over A–H ≠ Noul", "hourly 0947 / notes.md §119", "ggmlc GGUF is not llama.cpp", "serving substrate ≠ calibrated replica", "Qwen3.5-9B ≠ Archer", "planner writes JEV selects", "hourly 1049 / notes.md §120", "open recreation ≠ calibrated replica", "semantic lint is a sensor not a proof", "cutoff 0.8 still soft", "paired bootstrap CIs *theirs*", "Same accuracy, 35x faster *theirs*", "hourly 1143 / notes.md §121", "revisit HIGH / since-last-look", "catalogued repo changed", "star-noise vs material change", "densify prior notes without inventing equivalence", "decide is not generate", "tryDecide returns typed calibrated judgments not a token stream", "GLiNER/GLiClass ports are class members not Jev replicas", "93.5% *theirs* not Harbor", "74.9 *theirs* not Harbor", "8.7x *theirs* not Harbor", "Option-Marker joint attention", "openjev:0.2.1", "thinking=True/False per-field budget", "PLAN_Qwen35", "hyperspaceai/jevcache ≠ kushals256/jevcache", "wire-compat ≠ logit-equiv", "SHA move is not a replica", "hourly 1248 / notes.md §123", "typesafe-sdk 0.7 Pydantic response models", "msgspec dropped", "The server's output is unchanged and was never wrong", "SchemaError is 400 plain-string detail not 422 list", "Pydantic response models ≠ logit-equiv", "msgspec dropped is not a replica", "Error contract is not a Noul", "coverage-at-error-budget *theirs* not Harbor", "PLAN_Qwen35 still proposal for review", "GLiNER locate ports are class members not Jev replicas", "Locate ≠ decide", "~160 ms *theirs* not Harbor", "0.971 F1 *theirs* not Harbor", "hf:fr0stbit3/laya-gguf serving substrate ≠ calibrated replica", "jkcdarunday/SystemOne-Next ≠ TypeSafe System One", "hourly 1340 / notes.md §124", "vLLM NVIDIA + MLX Apple Silicon", "Codiv hosted free endpoint", "dual /v1/systemone + /v1/chat/completions", "chat 501 on MLX", "dual serving is not generate", "Hosted Codiv ≠ TypeSafe", "hr98w/jev-visual 167★ Apple Silicon visual candidate scoring", "37.30s → 2.40s at 64 decisions *theirs*", "Breakout 9 bricks 6 returns 2 lives *theirs*", "candidate probabilities are relative not correctness", "jkudish/jev-mcp 156★ ten MCP tools", "recommendation is advisory", "the server never blocks on its own", "TypeSafe CLERC 5% to 18% *theirs*", "jkudish/jev-mcp ≠ burnigtm/jev-mcp", "zhengxuyu/litjev off-the-shelf Qwen decision layer", "Probabilities are not calibrated by default", "Qwen/Qwen3.8-27B ≠ Archer", "zhengxuyu/litjev ≠ alexwestco/llm-to-jev", "Zefan-Cai/Open-Jev LoRA + scalar head", "2B 94.71% 9B 97.54% hard test *theirs*", "2B OOD 86.02% 9B OOD 91.97% *theirs*", "80,816 training rows", "27B still in progress", "LoRA ≠ RLCD replica", "Zefan-Cai/Open-Jev ≠ TheoLeeCJ/openjev ≠ razorback16/openjev", "cristianoliveira/jeq intelligence you can pipe", "pass-min 0.8 still soft", "JEQ does not own actions", "AndyInQtr/laya-coreai CoreML serving substrate ≠ calibrated replica", "AndyInQtr/laya-coreai ≠ mizorewww/laya-coreml", "hourly 1441 / notes.md §125", "TypeLLM/TypeLLM densify HEAD 6a48f9f1e623", "README densify 3k→12k B", "Batch 5.8x *theirs*", "Constrained AR ≠ calibrated Noul", "jaredpalmer/kev densify HEAD b339f446a0ef", "Kev-0.6B 4B 8B family", "4B new-source 0.790/0.806 *theirs*", "8B new-source 0.796/0.780 *theirs*", "Jev hosted 0.857 *theirs*", "Questions share the input text but cannot read each other", "No Jev outputs were used for training", "8.2% ≥0.9 on wrong *theirs*", "option order can change an answer", "Qwen3 ≠ Archer", "TheoOliveira/pi-jev 21★ fail-closed routing", "JEV_THRESHOLD 0.65 still soft", "harshwasan/jev-sentinel fail closed never auto-allows", "harshwasan/jev-sentinel ≠ leepokai/jev-guard", "jackbarunz/jev-tool-router ≠ esinocchi/jev-tool-router", "threshold 0.90 still soft", "76/81 vs 77/81 *theirs*", "0.419s vs 2.459s *theirs*", "$0.00486 vs $0.03673 *theirs*", "not a security boundary", "baronunread/leanest fail-open uncertainty means RUN", "classifier.dev default Jev/Laya pluggable", "openlayer-ai/jevals ≠ dayhaysoos/jevals", "estimates not Harbor", "classifier ≠ authorizer", "MrJev/awesome-jev 118 entries catalog ≠ endorsement", "MrJev/awesome-jev ≠ yibie/awesome-jev", "Koushik890/jev-firewall fail closed ask_below 0.7 still soft", "CompleteTech-LLC-AI-Research/jev-codex-approval experimental native not compiled", "confidence is not a measured probability", "rh-guard owns primary gates", "hf:rAVEUK/open-jev-deberta-v3-large encoder class member not Jev replica", "hf:p-yan/laya-quanto serving substrate ≠ calibrated replica", "hf:Gtrkrsk/laya serving substrate ≠ calibrated replica", "hourly 1542 / notes.md §126", "razorback16/openjev densify HEAD febf02e88989", "release 0.3.0", "re-pin vLLM PR #57250 restructured head", "MODEL_VERSION stays openjev-0.1", "uv.lock hygiene", "restructured vLLM head ≠ logit-equiv", "frostney/clean-code-review 7★ typed judgments not opinions", "documentation is read not judged", "morcoan/JMP Joint Model Participation", "Models participate. Real tools execute.", "Thresholds are policy not model", "Kelbie/hunch ≠ carldaws/hunch ≠ tpellet/hunch ≠ huncho", "Jev never generates prose JSX or code", "json-render is the only renderer", "game success ≠ calibrated Noul", "Shalimov04/open-jev ≠ razorback16/openjev", "MstyAI/laya-onnx empty repo", "hf:Praveenrajus/jev-bench HTTP 200 was 401", "hourly 1643 / notes.md §127", "TypeLLM/TypeLLM densify HEAD 702e6a287f3c", "truncated thinking then constrained decode", "0.8B thinking On 0/18 *theirs*", "forced closure 20/20 type-valid *theirs*", "jaredpalmer/kev densify live HEAD 8465c4c4c294", "Kev-0.8B completes family", "4B new-source 0.794/0.832 *theirs*", "9B new-source 0.812/0.837 *theirs*", "transfer-v9 Kev-9B 5% Jev 9% Kev-8B 26% *theirs*", "SemIf Kev-9B 0.917 Jev 0.965 *theirs*", "scienthoon 0.952/0.911 vs 0.897/0.914 *theirs*", "transformers >= 5.17", "Qwen3.5 ≠ Archer", "notque/vexjoy-agent 421★ /d routes /do fallback", "Facts go to code. Judgments go to Jev. Only facts can block.", "Jev never blocks", "jqueryscript/awesome-jev ≠ MrJev/awesome-jev ≠ yibie/awesome-jev", "five-lines threshold 0.80 still soft", "371ms $0.0000189 300-call *theirs*", "tpellet/jevify ≠ altryne/jevify", "seb4ez/jevguard-mcp ≠ seb4ez/jevguard", "resumocast/jev-mcp ≠ jkudish/jev-mcp", "Adrian-Ernesto/jevsort ≠ zzzzzec/jevsort", "MidasMulli/kev-ane 155/155 argmax *theirs*", "hourly 1746 / notes.md §128", "Fine-tuning on your own data", "--data JSONL", "--init_from warm-start LoRA/head PR #9", "from-scratch ≠ warm-start", "JSONL labels ≠ Harbor", "Kev-0.8B 4B 9B Qwen3.5 family", "0.33 vs 0.84 vs 0.83/0.88 *theirs*", "reconstruction ≠ replica", "assay-001 split verdict", "Promethe-us/awesome-jev ≠ MrJev/awesome-jev ≠ yibie/awesome-jev", "ThePFMind/jev-mcp ≠ jkudish/jev-mcp", "kyegomez/open-jev ≠ razorback16/openjev", "namenu/pi-jev-effort ≠ TheoOliveira/pi-jev", "samatv256/mini-Jev ≠ r-ms/mini-jev", "hourly 1843 / notes.md §129", "TypeSafe-compatible ≠ TypeSafe replica", "SystemOne.from_pretrained", "replica ≠ TypeSafe", "76.7% vs Jev 86.9% strict common subset *theirs*", "kotoba-lang/typed-decisions ≠ convaiinnovations/laya-typed-decisions", "DeBERTa-v3-large 0.855 / 42 ms *theirs*", "aisearchio 15-link census catalog ≠ endorsement", "user-provided 1936 / notes.md §130", "systems latency ≠ semantic equivalence", "hard acc ≠ calibrated Noul", "Open-Jev TREC pending", "GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*", "TREC-DL Jev/Luna/Astra completed", "customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*", "1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*", "prefix caching experimental/off by default", "not merged base models", "Open-Jev densify HEAD 4933ee84951f", "Astra TREC commit 1dd56990be7e", "densify §125 not a sibling first sighting", "Open-Jev densify / notes.md §125", "launch X thread https://x.com/Zefan_Cai/status/2101782158658695388", "2101786019607740436", "2101789698947793231", or "cascade sign-flip / calibration theater": read `references/faq.md`, + other", "training confronts Choice other / none-of-the-above", "soft AGENTS.md rules vs the linter", "screenshot Choice / omni System One", "extractive quotes / pointer not generator", "compaction summarize vs pointer", "encoder vs Jev compaction backend", "shadow-mode compaction rollout", "CI flaky-vs-real merge gate", "fail-open VOI wake/resume", "claim vs session evidence", "S1 indexer escalate-S2", "Harbor on/off routing", "fail-open vs fail-closed wake vs CI gate", "encoder vs Jev computer-use backend", "hybrid local decide + remote fill", "DONE vs verified success", "stdout prune vs session compaction", "OpenCode jev-pruner vs Claude jev-pruner", "zen-chat vs jev-zen Noul", "hard envelope then Noul prune", "Cua-S1 vs TypeSafe Jev", "plan vs execute dry-run", "specialist computer-use vs general agent", "local drop-in vs stub scorer", "route vs memory", "when does it hold / extractable from state", "decision model vs constrained LLM", "dual-process S1/S2", "combinatorial grid vs extractive", "uncalibrated local likelihoods", "decision-native RAG", "classify-first / read selectively", "living applied-mappings atlas / class patterns", "silence as safer / draft-gate heartbeat", "robotics text-state vs pixels", "verbatim ledger vs summary", "judgment as language primitive", "Stagehand extract pick-and-copy", "harness observe-score-act vs demo loop", "public judgment wall / six parallel questions", "meaning-search without embeddings", "attention ≠ correctness", "skills→oxlint / AST prove ∩ remainder", "session-sticky first-prompt routing", "measured RAG rerank vs generative rerank", "capability kernel / secrets never in the agent", "Jev is SENSOR not policy", "type-safe ≠ correct", "typed control plane around DSPy", "native vs verbalized confidence", "engine owns truth / Jev owns judgment", "human-confirmed kill gate", "train specialist vs few-shot hosted", "decide→policy→LLM leftover", "Noul 0.5 cannot-tell never rounded", "calibration ≠ sortable / ORDER BY", "pairwise inversion / Score ordinality / two-decimal ties", "wire-compat GLiFormer /v1/systemone", "class-backend economics", "loopback gateway hosted + local", "do not distill Jev as teacher", "active-learning triage", "evidence-packet explorer", "meaning-grep AND/OR/NOT", "closed-vote-only / no planner LLM", "Jev vs PCD Harbor", "PCD O(1) ≠ Noul", "host-owned handlers × System One", "OMP/pi fail-open gate", "permission vs probability / operator owns thresholds", "judgment ≠ permission / Jev never grants access", "eval integrity / instrument not score", "constrained optimizer + S1 features / never sole hot-path gate", "privilege ≠ verdict / effect contracts not tokens", "attention filter / VOI for human review / never blocks / never green unless sure", "measurement owns endorsement / evidence-gated question packs", "Jev supplies evidence / code owns authority", "ranking ≠ calibration / never hard-threshold raw p as frequency", "hot-click CU / indexed element table", "Jev judges relevance / code decides structure", "local rules first then remainder / never auto-train on own hides", "combinators / System One as control plane", "receipts not leaderboard / type-safe ≠ correct jaggedness", "VOI over skill library / skillranker abstention", "OOD calibration / AUC ≠ ECE", "Jev vs thinking-budget small models", "turnstile / replayable evidence≠authority", "MLX one-pass schema→JSON / Apple Silicon replica economics", "memory leases ended by new evidence", "never confidently wrong / TLA+ compose / escalate instead of hard-gate", "no seal no advance / coverage ledger / mint ≠ product brain", "skill-broker sibling / judgment ≠ permission", "sureness bands / max_prob is generous", "JevBench / calibration not in Main Score", "CI typed gate before expensive review", "Codex MCP host adapter", "judgment as attention redirect / jev-preflight", "compress-before-first-send / dizk jev-lens", "tools≠use / SessionStart over hoping", "observational memory / pi-om keep-kind", "open-Jev class / openvons / JevPick", "physical-world System One / HA-Jev / not for locks", "judgment outside the store / jevql", "landed-script trust / headless≠auto-approve", "digital-design combinators / extended five", "VOI cache admission / same-intent skip LLM", "BM25 vs Jev skill routing Harbor harness", "zeroshot vs BERT / contamination DiD", "typed escalate continue abort baton / inverted loop", "worth-your-attention VOI / ThinkyMiner Winnow", "Jev WHETHER Python HOW LLM WHAT", "conflict vs ignorance / named Choice escape", "Playwright executes Jev chooses", "OpenJev /v1/decide not drop-in", "SemIf wire-compat runoff; SemIf rename densify / MLX backend / 5.21× systems≠semantic / Softmax ≠ Noul (`notes.md` §117)", "decision-as-memory flywheel", "record/replay CI / jevassert", "failure-finding arena / jevarena ≠ jev-arena", "BBQ not a bias cert", "decider≠executor", "sentence-as-rule lint / jevlint", "sentence-as-rule lint / jev-lint is jevlint rename", "VOI hunk prune", "whole-repo intent VERIFIED/VIOLATION/UNKNOWN", "GLiNER2 spec ≠ replica", "open replica substrates / grande / laya-jolt / JEV-CPU", "ONNX local-jev not equivalent", "persist constraints across compaction / pi-heed", "calibration+cost first-class gates", "Harbor-shaped Jev vs SGR LLM-as-judge / jev-judge-bench ≠ jevarena ≠ jevbench", "hand no-text steps / jev-use / Vercel drops confidence", "Pi System-One control plane / pi-jev-control", "never free-generates / jev-gpt tree of Choices", "OpenRouter recipe atlas / samples not benches", "personal history feed / jevfeed / no social graph", "competing NAR claims / dual-channel ECE / openJev-verdict ≠ OpenJev", "empty compaction-proxy skip / IPECTER", "throughput ≠ latency / like-for-like ECE", "1-token logprob endpoint ≠ Noul / coverage ≠ correctness", "open replica engine / jevinf", "unofficial Elixir SDK ≠ OTP peer", "jevex n=16 files-to-read VOI", "commit pre-review attention≠verdict / middle band", "Hermes plugin is Agnes not TypeSafe", "pi-jev-compact ≠ pi-jev-compaction", "empty Codex-proxy skip / IPECTER runway", "decision-native inbox / mailordinal", "unofficial jev-cli not ready / ≠ jevql", "laya-multilingual / English checkpoint confident-wrong OOD", "schema-scorer peaked ranking ≠ calibration", "HF 401 / GitHub 404 Hub-only", "productized System One HTTP / classifier.dev", "escalate-under-threshold / smart tier / multi-label ignores", "silent FALLBACK / granite 0.546 vs advertised 0.800", "vs_jev tracked JSON / read eval/README", "choxos/jev-reviewer ≠ egma-ai / systematic-review pointer", "two-pass Choice+Noul evidence extraction", "not-found is an answer", "human check as productized judgment", "githubnext/localjev ≠ kunchenguid/local-jev", "wire-compat ≠ logit-equiv / prompted JSON ≠ structured read", "self-reported probs / entropy confidence", "GitHub Next local /v1/systemone", "LM Studio runner gap / structured-read primitives", "NandhaKishorM/laya packaging ≠ Hub-only / Router script-before-p", "post-T ECE ≠ raw ECE / Banking77 token-budget", "0.85 still soft / not TypeSafe drop-in", "external census ≠ scored bake-off", "GLiNER2+routers class-boundary", "incomplete openjev census vs watch", "Harbor honesty watch / silent fallback", "JevBench v1.2 geometric mean / cal ON rank / weight sensitivity", "option-order 72→21 / instruction models in the class table", "self-host latency ×2 assumption / est. costs", "Laya absent is a gap not a named exclusion", "Qwen3.8 27B ≠ Archer", "hourly already-folded watch / apply-the-five / skip thin noise", "hard-gate Noul as PR/quality gate is soundness theater", "S1 never stalls waiting / S2 one-use advisory", "Local controller ≠ githubnext/localjev", "purple telemetry = consumed not arrived", "seed = geometry not async replay", "20% starting gate still soft", "no pixels to either provider", "OCR+AX observe-score-act / typesafe-computer-use", "never send screenshot to frontier for the decision", "overlapping CU options = false low confidence", "split kind/item/site", "155× one-screenshot ≠ Harbor taskset", "decision ≠ answer-reader capture", "ASR observe-score-act / jev-voice-browser", "partial-speech VOI / free-text waits", "spoken confirm ≠ hard auth", "numbered overlay without another model", "wrap-as-execution / AgentGhost ALLOW ASK DENY", "rules first then Jev remainder / ASK throws / fail-closed", "reddpy/AgentGhost ≠ jwen5419807/agentghost ≠ vventirozos", "JP genre atlas / studio_yebisu / stars ephemeral ≠ eval", "Jev Clearly Explained / akshay_pachaar / LLM hammer", "schema-safe ≠ correct / 200× 400× TypeSafe ceiling", "questions-as-code / shadow first / not a TypeSafe how-to", "proposition ≠ embedding / contrast-set", "boolean composition of soft Nouls / AND OR NOT", "uehaj/jev-semgrep ≠ semgrep.dev", "meaning-grep dedicated fold / not a gate", "decision-validated UI / Jev never authors text", "decision-as-assert / jevtest ambiguous band", "typed decisions drive UI / jev2ui", "hybrid S1 closed verb menu / anima3", "pointer-not-generator search / JevFind", "jev-frontier-bench ≠ frontier-100", "product bakeoff ≠ architecture duel / GLiClass", "four engines same questions / majority floor", "authorship named escape / not evidence", "ha-switchboard HA remains execution", "n8n classify/route/score / Low Confidence", "fast-jev-compaction-pi ≠ pi-jev-compact ≠ pi-jev-compaction", "jevloop full-distribution optimizer / no LLM in the loop", "laya-vision SmolVLM / score untrained", "Cerebellum-2B /v1/decide ≠ TypeSafe / wire-compat vs agent-routing", "laya-grounded not drop-in / Platt not temperature", "GestaltLabs/Jeff-1 ≠ logan-markewich/jeff / acc vs ECE n=9730", "stanley-code empty findings ≠ approval / human promote", "findme ≠ JevFind / NL memory beam-search FS", "jevsubrouter price workers not conversation / counts ≠ dollars", "feelings .feels() default 0.5 is Noul-0.5-never-rounded / ≠ hunch ≠ Probably", "apa-agent-harness ≠ AntonioCoppe/jev-harness / unpublished npm", "grok-bot-jev skill cannot force a bot that ignores it / A/B proxies not tokens", "Essentiel-Jev never authority / human every action", "enzo-mcp independently falsifiable claims / ≠ jev-sift", "pigeonhole OTHER skip / decision-as-filing", "jev-reliability Nothing about accuracy", "clduab11/jev-test ≠ realZachi/jevtest / Nothing runs yet", "jev-rag-benchmark Jev wins is not an assumption", "dairui1/jev-lab ≠ BrendanH18/jev-lab", "jevmail gmail.readonly / mailjay archive/trash", "ZHUBoer/ego-jev reserved __none__", "runWorkflow completed ≠ success", "jsort scores are relative", "Noul not Choice for scale", "groundedness-judge-bench native vs schema-guided", "implicit_true included in yes", "jev_playground 0 promotions", "routing-backtest 0.0447%", "yuyang2230/jev-agent-skill jev-1.13-free", "jev-techstack-classifier stack_config.json", "s1_ruby collapse late", "undecided? abstain", "2389-research/judgement license null", "confidence ≠ winner p", "typesafeai-sdk-community not a new species", "tpellet/hunch exit 3", "never-execute list", "jev-file-search scores not calibrated accuracy", "jev-linkmap Jev never sees S2 prose", "muhammedilyasy/jev-mail metadata only", "tidy none-of-folders stay", "tab-bouncer pinned/audio/current never closed", "lkclean Show fail-open", "jev-yt-time-saver Show anyway", "ORIGIN pause-if-no-Jev", "validResponse sums-to-1", "jev-crawlers risk bands never raw boolean", "jevbrain AUTO_ACT is not a Noul", "judgekit YAML classify/score/route/verify", "typed-judge-kit verdict-in-code", "alsoleg89/decide packing VOI", "0.8 ≠ 80% accuracy", "Jev-Calibration Platt ECE 0.117→0.052", "jev-calibration-arena never acts", "ctmx/openrouter-jev-mcp Decision-as-Plugin", "FrancoisChastel/jev-code ≠ npm jev-code", "claudecode-jev-marketplace fail-open not hot path", "pedroknigge/mcp_jev packs not ask_jev", "cyrusasco/typesafe-mcp noul deadband 0.35–0.65", "codaaiteam/jev-skill jevtypesafeai.com ≠ TypeSafe", "hermes-switchyard ≠ hermes-jev-router ≠ hermes-plugin-jev", "nanoprune 2.8MB ECE 2.58%", "smartdio/jev-browser-agent ≠ ZHUBoer/ego-jev", "Dakai/omp-jev-web DONE ≠ proof", "hari007sh/jev ≠ dannote/jev", "0thernet/system-one-skills deterministic verify", "typed-gate band [0.40,0.60] is refusal", "pi-jev-gate fail-closed; choice is the verdict", "Foq ~25ms/2.2GB local", "rev prefill-only + HF jev-0.5b", "robfrase/jev planning memo", "typesafe_agent_gates 27/27 / 31/31", "EpicEric/safe-sh static remainder", "pastepilot Confirm before act", "Jev-Reranker live Jev not yet measured", "sessionwise opt-in relevance", "jev-search pointer sieve", "400ms Salesforce WebMCP", "typesafe-scheduler-diagnostics advisory", "droidjev screenshot-free", "Tewoto1 jevcu planner still writes", "ha-conversation-jev Jev→Grok", "dsh-jev can only gate", "jev-classification-benchmark specified not run", "jev-luna-pagerduty p≥0.50", "meldltd/meldecision laya-go ONNX", "laya-doom never pixels", "logixism/laya-api empty README", "akpsahan/laya ≠ Archer", "choxos/jevchess engine owns truth", "jev-drive sim not AV", "story-arc Jev never authors", "jev-hs-assistant HS6", "golergka/jev-plays-starcraft-2 UI-verified ≠ API Victory", "awesome-jev-use-cases catalog", "Nibir1/typesafe-go ≠ official", "fingerprint after redact", "recall vs decide", "publish fingerprints+answers", "CI replay as Harbor cousin", "Cache hit ≠ correctness", "hyperspaceai/jevcache ≠ kushals256/jevcache", "human labels only", "score never auto-accepts", "production capture flywheel", "sutro-sh/jev-align ≠ caiovicentino/jev-align", "guidance ≠ hook", "catalysts ≠ summaries", "compile-time System One", "unofficial ≠ TypeSafe", "format_version modernbert-jev/1", "Argos1111/jev_local ≠ us/jev-local ≠ kunchenguid/local-jev", "LFM default ≠ ModernBERT backend", "Nemotron ≠ TypeSafe Jev", "not a calibrated replacement", "djev-dev complements djev-spark", "images as Choice options", "Laya essay numbers *theirs*", "Router/OOD confidence", "hosted bootstrap ≠ silent TypeSafe", "difficulty + policy thresholds + JSONL trace", "jev-codex-pilot model + reasoning depth", "keep/shadow/hybrid/reject", "quarry evidence projection", "Frank-ZY-Dou/awesome-jev robotics/3D/control", "one-dollar-tahoe TypeSafe Jev defense eval", "jevguard calibrator/cache/escape", "jev-ci-selector CI shadow mode", "llama-jev llama.cpp replica", "petercr/jev-orchestrator ≠ FleeexCorp/jev-orchestrator", "seb4ez/jevguard ≠ AseemPrasad/JevGuard ≠ pablozr/JevGuard", "webNeat/llama-jev ≠ WiktorB2004/llama-index-jev", "OpenCode jev-pruner context sieve", "observe→score-candidates→prune", "jev-zen / jev-1.13-free", "zen-chat ≠ Noul", "fail-open original", "keepScore >0.1 floor", "host port of tamaratran/jev-pruner", "indiejoseph/opencode-jev-pruner ≠ nrdz-labs/fast-jev-opencode", "jev-webagent-bench empty stub", "Kiln-AI/jev_jsonschema noul_threshold 0.5", "NSStudent/JevSwiftSDK unofficial", "GLiNER2 native Apple path", "unofficial Swift/Core ML GLiNER 2.5-small", "entity spans + confidence", "not Choice/Score/Noul", "not TypeSafe", "label descriptions as schema", "on-device ANE economics", "honesty locks", "shershah1024/gliner-native-runtime ≠ Fastino", "≠ gliner25-compaction ≠ gliner2-ultrafast ≠ Eran-BA/Jev_from_GLiNER2 ≠ NSStudent/JevSwiftSDK ≠ jevmlx", "default threshold 0.1 still soft", "soft Noul ≠ hard safety", "Decision Graph Protocol frame→assess→commit", "app retains permissions/effects", "Jev-first assessor-neutral", "guarded commit / receipt/next frame", "assessment batching", "hard-gating DGP as safety theater", "numerous-com/dgp ≠ TypeSafe official", "jegrep calibrated path+range Nouls", "no embeddings/index/daemon", "~$0.01–0.03 typical", "agent --json", "can1357/jegrep ≠ Bentlybro/jevgrep ≠ uehaj/jev-semgrep", "Archer-arch fidelity", "kev family OOD 0.76–0.77 vs Jev 0.86", "block-causal isolation", "pointer/readout CE-trained", "/v1/systemone drop-in", "replica honesty", "cost-sensitive decision theory × System One probabilities → control flow", "thresholds derived from costs not hard-coded", "YES / NO / UNSURE from cost_false_yes / cost_false_no / cost_human", "auto-batching same-object questions", "Kungie/gut ≠ tpellet/hunch ≠ carldaws/hunch", "judgment vs generation", "deterministic execution after probabilistic judgment", "exactly one app-owned callback", "explicit uncertain branch", "Illusion47586/judge ≠ lexingtonhibiki/judgekit ≠ Ascurse/typed-judge-kit", "variable-N option scoring as the trainable object", "dynamic candidate bags not fixed label sets", "zwliJay/jev-forge ≠ NanoJev", "open replica economics / latency vs closed Jev", "NAR local drop-in", "wfzyx/von late-catch HIGH", "competing NAR claims / replica honesty", "typed judgments vs chat judges on guardrailing", "ishaannk/llm-vs-jev cross-note only", "deeper integrity fold is rh-guard", "nothing wins outright", "can be argued out of guarding"", "Jev IS the if-statement", "judgments/probabilities drive branches", "text model only writes prose", "interpreter owns variables/loops/budgets/replay", "otherwise maybe / confidence gate", "chaos samples after the gate", "southpolesteve/probably ≠ carldaws/hunch ≠ feelings ≠ Kungie/gut ≠ Illusion47586/judge ≠ tidymodels/probably", "133★ / forks 10 live", "build calibrated classifiers from human feedback", "retrieve by relevance not resemblance", "one calibrated yes/no per memory in one request", "pointer mode 17/18 19/20 *theirs*", "embedding resemblance misses the allergy", "samdotmak/jev-recall ≠ jev-search ≠ jev-sift ≠ carryforward ≠ chopratejas/invalidate", "memory leases ended by new evidence", "six Nouls then fixed rules in code", "0 of 157 false invalidations", "questions/plans/directives are not evidence", "unsure → review queue", "host keeps the store", "name↔body / comment truth / test-claims", "mizchi/jev-lint is mizchi/jevlint rename", "no shipped rule has severity error", "~1 in 5 findings wrong *theirs*", "mizchi/jev-lint ≠ huntedman/JevLint ≠ MichitoSugawara/jev-lint", "JSON Schema → typed JSON via Jev", "noul_threshold 0.5 decoder not a proof", "IncompatibleSchemaError lists every bad property", "on-device Laya CoreML ANE", "~5 ms P50 short decisions", "189/189 FP16 checkpoint parity", "10× not achieved", "mizorewww/laya-coreml ≠ gliner-native-runtime ≠ jevmlx ≠ NandhaKishorM/laya", "softmax over allowed tokens ≠ Noul", "question-first cache", "Micha0827/snapjudge ≠ githubnext/localjev ≠ jevmlx ≠ cendress/SnapJudge", "Jev-first Pi agent loop", "slow-LLM fallback", "explicit action menu / CandidateSource unimplemented", "62 tests wiring not quality", "direwolfiy/JevPi ≠ standardagents/jevpilot ≠ pi-jev-control", "resume-screening bias audit methodology", "name×resume factorial independent Nouls", "callback determined by resume quality", "mean-probability name gaps operationally negligible", "natemoo-re/bias-bench ≠ BBQ", "Plan/PRD panel → code-owned pass|review|block", "cheerleading out of scope", "austindixson/planalyzer ≠ single-goodness Noul", "cost-aware multi-model routing/escalation", "decide vs do", "successful-task cost", "cannacre8ive/switchboard-ai ≠ ha-switchboard ≠ hermes-switchyard", "frozen-protocol zero-shot bench", "TypeSafe Jev vs PrismNLI vs Laya", "contamination caveat", "elcronos/jev-vs-open-decision-models ≠ JevBench ≠ DMB", "context-window admission control", "VOI gate which tokens are worth the expensive model", "fail polarity per lens", "on small inputs lenses lose money", "cvsgireesh/jevusher ≠ jev-sift ≠ winnow", "typed decision control plane", "receipt ≠ authorization", "historical-v0 zero retained cases", "MokiMeow/jev-fabric ≠ jev-forge ≠ dgp", "live 15-dim typed rubric re-score per pause", "scoring economics exemplar", "OpenJev/Codiv ≠ TypeSafe hosted", "jose-troche/live-rubric ~$0.000004 desc / ~$0.000006 README", "adversarial pre-registered Jev eval", "28 predictions before data", "123,805 requests", "confidence does not track ignorance", "polite injection 65% / crude 0%", "willkelly/jev-evaluation ≠ jevals ≠ jev-baselines-eval", "provider-neutral Elixir/BEAM Noul/Choice/Score SDK", "class infrastructure", "nshkrdotcom/system_one_sdk ≠ typesafe_sdk ≠ dannote/jev", "question-linting of Jev questions themselves", "nine jaggedness rules, no API key, no labelled data", "static lint ≠ measured separation", "yodablocks/jevq ≠ tenbin ≠ JevLint ≠ commitjev", "open-weights Laya as class exemplar (binding)", "Nx/Bumblebee runtime", "host chooses backend", "ChristianAlexander/laya_ex ≠ system_one_sdk ≠ dannote/jev ≠ NandhaKishorM/laya", "on-chain/edge Laya deploy", "parity_verified stays false", "model output never grants Tx", "humandebri/IC-Laya ≠ laya_ex", "auditable weekend replica", "Jev outputs never used for training", "soft human-vote distributions", "unpaired 0.577 vs 0.727", "agilabs-ai/jev48 ≠ JevBench ≠ Mapika/decider", "adversarial dual-judge / framing attack surface", "comparative framing is the usable judgment", "prior injection crowds out evidence", "copyleftdev/ember ≠ ember.js", "Laya specialist fine-tune pipeline", "training still GPU-pending", "PIXELZX0/XERON ≠ convaiinnovations/laya", "Hub Laya replica drop", "daliborsb/laya ≠ convaiinnovations/laya ≠ NandhaKishorM/laya", "System One student distillation corpus", "gold is programmatic", "teacher is closed-API clone", "do not distill Jev as teacher of record", "MagaBitmex/jev-4b-distill-data ≠ missing student checkpoint", "non-LLM VIN System One", "planning depth not chat", "lewislululu/jevon ≠ douglance/jevon", "source-bound evidence checks", "local quote mismatch needs no API", "exit 0 ≠ claim truth", "WaynezProg/jev-kit ≠ jonathanavis96/jev-kit (Airlock) ≠ jev-use ≠ jev-mcp", "independent System One evidence catalog", "scores not one leaderboard", "no external record currently reproduced", "TokenTrim no-Jev matched hybrid 62.4%", "reachjalil/system-one-bench ≠ mallahyari/system-one-benchmark", "21 tasks · 134 items · 208 questions", "scenes from public GitHub contracts, not production logs", "SivletLabs/jev-eval ≠ willkelly/jev-evaluation ≠ 4esv/jev-eval ≠ xxkuboxx/jev-eval ≠ onlyoneaman/jev-eval ≠ dayhaysoos/jevals", "option isolation (sibling-blind)", "permutation-equivariant", "Hub OWNER not published", "nafisazizir/hev ≠ jaredpalmer/kev", "frozen local LLM logits, no trained decision head", "residual-head 9,222-param decreased 73/96→67/96", "confidence = 1−normalized entropy, not P(correct)", "yuki-oshio/mini-jev ≠ r-ms/mini-jev", "Jev classifier as autoregressive next-token predictor", "ChatJev-style soundness theater", "erik-dunteman/ChatJev ≠ dannote/jev ≠ jev-gpt", "calibrated decision head × AlphaProof value head", "implementation-layer isomorphism, semantic difference", "timeout = censoring", "do not launder Noul as proof", "parallel rank-prediction vs serial selection", "independent questions can conflict", "zzzzzec/jevsort ≠ keltokhy/jsort", "curated open System One ecosystem catalog", "rupeshpoojary9/awesome-open-system-one ≠ AnotiaWang/awesome-jev", "arXiv paper radar with Jev relevance scoring", "ranking ≠ calibration / 0.5 still soft", "fail-open failed evals not marked seen", "train calibrated ~27M from scratch", "typed Q→prob dist / one forward pass / no LLM decode", "hyusi2003/MiniSystemOne ≠ Colvin0315/MiniSystemOne", "description-only stub / size 5", "ESCI hard probe fails four of six", "jev_bool ECE 0.242 inversion 0.255", "do not re-fold §60 six-gates as new", "jobbyjev one-request-per-company from batch-size result", "find/design/evaluate TypeSafe Jev decision loops", "karanb192/jev-architect ≠ samtay32/jev-system-architect", "Jairik/jev-distiller size 1", "distill-Jev UI stub / do not distill Jev as teacher of record", "post-launch scored use-case map / Jev self-scores then human curation", "licensedsaucer9-web/jev-opportunities", "Jev-inize a use case into classifier/router", "gavinHuang/jevinize → simple-jev not TypeSafe", "featherless-ai/simple-jev", "compare saved decisions / same label can still change the branch", "VihaanAgarwal/jev-diff ≠ Saik0s/diffusiongemma-jev-macos", "not tested with a live Jev API key", "constrained logprob + temp/Platt ≠ Noul", "OpenJevPro pastes openjev-sglang JevBench as own", "zhangcy122/OpenJevPro ≠ IamBusy/OpenJev ≠ ekzhang/openjev-sglang", "PolyForm Noncommercial", "SmolLM-135M / sub-70ms / 0 output tokens", "demo P(True) 0.5052 / Choice conf 0.2872 / Score conf 0.0055", "README claims MIT / GitHub license null / no LICENSE file", "patelvishwa112/jev-system-one-rlcd ≠ arnabgho/rlcd-lite ≠ blackwood-rlcd", "source-backed Awesome Jev radar / 306+ commit-pinned", "logicrw/awesome-jev-projects ≠ AnotiaWang/awesome-jev ≠ yibie/awesome-jev ≠ cobanov/awesome-jev ≠ rupeshpoojary9/awesome-open-system-one", "auto GitHub sync / Issue-only submissions", "hashed n-gram encoder / rival-aware attention", "olanotolu/jevbetter vs jevlike starter", "synthetic hard menus top-1 0.916 vs 0.873 / ECE 0.0182 vs 0.0367 / 40 vs 4608 menus/sec", "shuffled-context control 0.335", "Turn any open LLM into System-One Jev", "uspraveen/Jevify ≠ Mintzs/jevify ≠ gulagala001/jevify", "Jevify-any-LLM architecture probe", "description-only stub / size 0", "Train encoder-only calibrated decision models from a task sentence", "Exu is a toolkit, not a method", "strictly proper scoring rule", "Pre-alpha", "Ruivalim/exu-base", "scratch-trained calibrated decision model", "typed Q → probability dists", "Colvin0315/MiniSystemOne ≠ hyusi2003/MiniSystemOne", "no published weights download URL", "90.5 seconds / 29.2% pipeline evidence", "p_i/p_j independent of other candidates", "Recipe for calibrated decision models — small model out", "init → synth → train → eval → serve", "91.1 % / ECE 0.022 *theirs*", "Jev zero-shot 75.1", "scienthoon/luce", "Put Jev's three headline claims on trial", "0.5B local GPU", "46x speedup / accuracy identical", "ECE 0.624 sentiment catastrophe", "bigger model worse calibration", "RichardoMrMu/jev-mini ≠ yuki-oshio/mini-jev ≠ r-ms/mini-jev", "System-1 decision engine for local LLMs", "structured choices only", "JSON parse of generated text ≠ Noul", "TypefAI JEV / Journal Entry Voucher", "tapsin/jev-local ≠ us/jev-local ≠ Argos1111/jev_local", "Jev 1.13 reward-model eval across 8 benchmark tracks", "40,940 examples / 0 API errors", "RewardBench v1 92.58%", "Precise IF 50.63%", "goya4140/jev-reward-model-evaluation", "Scaffolding in progress", "Jev vs LLM support-ticket routing", "static + live decision bench", "TypeSafe's own published benchmark", "illustrative simulations, not live API calls", "JevBench v1 — smart/cheap/fast/reliable", "I/C/S/K 25% geometric mean", "classifier.dev fast tier 84.8 is Jev behind its own API", "do not re-fold §78 v1.2 board as new", "Laya (421M) 70.1 now on board", "Zero-shot/few-shot LLM routing", "hard budget filter before Jev", "Jev never asked to perform budget arithmetic", "Jev judges the next state, XState enforces transitions", "simulation uses synthetic keyword fixtures", "catalog gravity", "v-modal/awesome-jev-tools", "★339 live REST", "curation is not endorsement", "crawler-maintained directory", "Daily GitHub + npm sweep, human-merged", "RadRebelSam/awesome-jev ≠ AnotiaWang ≠ yibie ≠ cobanov ≠ logicrw ≠ v-modal", "HF peft SPLADE/BGE reranker", "rdxtremity/jev-reranking ≠ carlaiau/jev-reranking", "query-side encoders, not a Jev replica", "ONNX System One Qwen3.5-4B scorer", "source:pngwn/system-one-qwen3.5-4b-scorer", "CC-BY-NC-4.0", "temperature 1.75", "transformers.js AutoModel cannot load this graph", "Consistency benchmark Space", "This Space contains no benchmark result yet", "12-case plumbing fixture", "Benchmark-driven Jev router and judge", "cheap alone is not success", "Jev does not write, sum prices, or claim accuracy %", "Sol 94.2 / Luna 83.9 / Jev path 89.7", "19.2% Sol / 62.3% cost save / 4.5pp miss of 2pp non-inferiority", "p50 latency worse than Sol due to routing overhead", "erendikmenn/jev-llm-router-benchmark ≠ jev-rag-benchmark ≠ ryantsai/jev-llm-router", "Express + node:sqlite", "mock and Jev decision engines", "previous_ticket_count >= 3 is code", "MIN_CONFIDENCE 0.6 still soft", "substring false positives", "aesaganda/jev-ticket-router ≠ SarathChandraBellam/jev-vs-llm-ticket-router", "Universal Figure & Diagram Router", "confidence ≥ 0.85 hard-gate is theater", "generative AI banned from scientific plots", "six visual branches", "hoangngochuong24947-gif/jev-figure-router", "human-labeled (state, question, label)", "166,054 rows / 22 configs", "soft_label for human uncertainty", "Praveenrajus/jev-bench ≠ fstandhartinger/jevbench", "ternary bonsai System One GGUF", "openjev's mechanism, Bonsai's weights", "Hub does not ship weights", "100/100 easy T/F is not Harbor", "label_mass ≠ correctness", "stock llama.cpp Q2_0 silently gibberish", "NicolaiMTLassen/open-bonzi-jev ≠ NicolaiLassen", "transformers.js DeBERTa ONNX", "source:com-kotobalabs/open-jev-deberta-v3-large", "temperature 1.05", "AutoModel from_pretrained works", "onnx-community/open-jev-deberta-v3-large-ONNX ≠ system-one-qwen3.5-4b-scorer-ONNX", "107★ densify", "GH 151M vs README 149.6M", "PR #1 now closed unmerged", "do not re-fold §71 claim-audit as a beat", "typed decisions, RLCD, confidence-gated routing", "structured ≠ correct", "mock not live API", "26 tests", "wjdjdakf17/jev-study ≠ baekenough/jev-study", "bonzi-27b-v2 / ternary-8b / 27b-v1 GGUF family densify", "WANLI-256 74.6% / 65.2% / 71.1% *theirs*", "Bonsai 1 27B Q1_0 runs on stock llama.cpp", "ternary still needs PrismML fork", "hf:heman10x/openJev-verdict-2.0 twin tokenizer-only", "OpenJev Vision image classification + uncertainty", "CLEVR-4 held-out joint 0%", "hfdataset:IamBusy/OpenJev-Vision-Research-v0.1 12,832", "294,912 derived targets not independent samples", "Laya multilingual ONNX WebGPU typed-decisions port", "63/63 selected answers / 5.1e-4 CPU / 1.2e-2 WebGPU", "UpHash-Network/mini-jev is yuki-oshio transfer", "jev-injection-bench 11,900 labelled prompts", "Jev best ranking / Haiku better ECE 0.021 vs 0.058", "0.5–0.9 band is where Jev's numbers do not mean what they say", "Prompt wording moves panic 28%", "manojlds/jev-dspy-bench ≠ dspachos/jev-dspy ≠ jmanhype/jev-dspy-lab", "Jev agreement is similarity, never ground truth", "no aggregate quality grade or merge gate", "AbstentionBench-on-Jev rank 1 of 20 vs 2025 field", "question-asymmetry", "forward-looking 0.465 never extreme", "openkev calibration layer not a runtime", "ECE vs coverage independent", "select_threshold returns inf", "escalation catches uncertainty not ignorance", "misakaikato/openkev ≠ jaredpalmer/kev", "pdf-race Docling→Jev vs Gemini", "parser owns the wall clock", "12/12 tie is a tie", "titles selected not generated", "flopcheck 16 calibrated tweet judgments", "mechanical tells in code", "ZeroX-01/jev-atlas ≠ Zaious/jev-capability-atlas ≠ gorock007/jev-atlas", "Laya calibration lab Gradio MCP", "T never changes argmax", "confidence ≠ top-label p", "easy probe set refused", "40–48 rows too small to ship T", "Gemma-4 26B-A4B jevify classification+calibration", "LoRA adapter twin not independent eval", "Gemma-4 E4B jevify", "E4B LoRA stub card", "kushalpatil/jevify-gemma4 ≠ Mintzs/jevify ≠ gulagala001/jevify ≠ uspraveen/Jevify", "GH kushalpatil07/jevify 404", "PAWS 0.580/ece 0.288 is the weak cell", "smaller E4B slightly better OOD ECE than 26B-A4B", "Hub jevify merged LoRA ships weights", "bonzi Bonsai-8B v1 GGUF densify", "Bonsai-1.7B v1", "Bonsai-4B v1", "WANLI-256 64.5% / 60.2% / 52.0% *theirs*", "rank #4 / #5 / #6 of 6", "JulesHuisman/jev-eval scaffolding / README SHA c356a584 (was empty e69de29b)", "JulesHuisman/jev-eval ≠ SivletLabs/jev-eval ≠ willkelly/jev-evaluation ≠ 4esv ≠ xxkuboxx ≠ onlyoneaman ≠ dayhaysoos/jevals", "7 bands 6/10 vs 40 bands 0/10", "source receipts + confidence slider re-policy without re-inference", "32/32 synthetic is smoke not production", "classify HF datasets across typed semantic dimensions", "roadus2 watch misspelling; lock roadius2/ultra_laya", "ultra_laya REVIEW defects", "default branch claude/laya-jev-review-gg5ppo", "XNLI EN 88.3% ECE 0.032 → RU 77.3% ECE 0.096", "Δ −11.0 pp [−14.2,−7.8]; ECE +0.063", "MASSIVE no detectable difference at n=600", "confidence is function of p_max (r=1.000)", "pointer-not-generator 400 human-authored responses", "proposed ≠ authorized", "FewRel 160: Jev 85.0% vs lexical 13.125%", "gated 100% (95/95) coverage 59.375%", "J++ composable semantic computation language", "judge-jev 0.5 still soft", "947 repos scored; A 273 / B 302 / C 372", "LLM rubric ≠ benches", "No benchmark winner is claimed", "phishing: naive 62.6% vs regex 91.8%; 5-atomic + LR 95.0% *theirs*", "AITuber tension ±15", "README npm global; repo is Rust", "git-confess code owns counting/blame/ratio", "httpx exhibit 11% (13/119) *theirs*", "90d trend +12.40% vs random +12.75% vs BH +41.71%", "5m win rate 25%", "Awesomejev 656 entries / 38,160 stars", "tracker likes 64 (+4) lastModified UNCHANGED", "Laya present; Blackwood ABSENT; Archer still promised_not_landed", "Blackwood tracker ABSENT; likes 2 gated manual", "r = c - p_a", "ECE 0.021; acc 0.807 vs warmup 0.746", "Independent primitive", "11.57s vs 54.10s · 4.67× · 120/128 *theirs*", "default path is pretrained Gemma probs not trained RLCD head", "GH Meanblock 404; lock leesk212/JEV-CPU", "softmax over letter slots ≠ Noul", "WANLI 0.741 vs openjev v2 0.77 *theirs*", "3-way NLI ≠ Noul", "priority 0.464 = majority floor", "banking77 contaminated", "raw margins not probabilities", "do not distill Jev as teacher of record (they distilled Haiku)", "“0.9 is not one number”", "ranking ≠ calibration", "banking77 0.8–0.9 stated 0.86 actual 0.73 over-confident *theirs*", "≠ Praveenrajus/jev-bench ≠ fstandhartinger/jevbench", "$0.0000153–$0.0000226 vs circulating $0.0004 (~20×)", "Score is 0..n-1 expectation not 0–1", "Noul has no confidence field", "TCP floor 198.8 ms", "type reliability is not a reason to choose Jev (json_schema 5/5)", "gateway tax not one number", "Function-only 5/8 vs hybrid 8/8", "4/8 without Jev", "8 designed cases not conversion lift", "200-row pilot Jev 86.5% 173/200 vs Gemini Flash-Lite 86.0% 172/200 vs Pro 87.0% 174/200 *theirs*", "not a ranking", "情緒測謊器", "8-example Jev vs GPT-5.6 Sol ~64× cost 5.4× latency *theirs*", "synthetic; no inference", "≠ JevBench v1.2 §78", "Judged 3317 / listed 2560", "Jev judges, code applies policy", "APA “microsecond policy / zero hallucination” overclaim", "Client-side quiz; pointer from held docs; scanned-PDF warn", "Jev judges / agent reasons / user decides", "selecting an option is not permission to implement", "pattern exact, judgement must clear floor", "no matching pattern → no model call", "not a correctness oracle", "Spec vs artifact remainder", "treating 0.85 as 85% / minProbability hard-gate as Harbor", "VERIFY acquires discriminating evidence, never same-pool confidence-only rescoring", "fast/full/max are ceilings not sizes", "Solar writes, Jev chooses NEXT ACTION", "do not reopen or amend PR #23 or #24 or #25 or #26 or #27", , "Calibration is not alpha", "NO CURRENT ALPHA CANDIDATE", "ΔR² approximately +0.00084", "Brier 0.2131387", "ECE 0.0421875", "Adding Jev probability to deterministic volatility improved Brier by only 1.4058e-05", "default 0.5 keeps zero non pinned", "keepResult median 0.14 to 0.17", "keepCall median 0.28 to 0.35", "usable range is about 0.10 to 0.25", "7.8% to 57.9%", "judges results it never sees", "task-finish eval not built yet", "$0.002 per compaction", "slavadubrov/sgr-judge-bench ≠ slavadubrov/jev-judge-bench", "Jev 108/120 $0.083 0.34 s", "Luna SGR 114/120", "paired Jev accuracy-difference intervals include zero", "not evidence of equivalence", "GLM SGR 26/120 93 format failures", "Terra-planned Jev hybrid 55/120", "rule-based by default, optionally Jev-backed", "empty README", "missing key cannot break the experience", "prefill plus exactly one decode", "softmax over A/B/C ≠ Noul", "BBQ 9,053/10,000 (90.53%)", "ECE 0.0890", "Mean confidence 0.9943", "overconfident", "score and noul not implemented", "DGUI 12 rows (was 6)", "INSTRUCT 119 rows likes 2", "encode the state once, decide everything in parallel", "0.740 accuracy against a 0.508 majority", "ECE 0.047", "fine-tune's advantage ends where its 384-token training data does", "jasonkneen/open-jev ≠ pngwn/open-jev", "same sha d41dc3cd", "Space does not call Jev", "recomputes routing from saved probabilities", "200-case Jev 97.0% / 100.0% / 95.0% / MAE 9.22", "synthetic repository benchmark", "Jev evaluations are advisory", "YehuiTang0316/jev-nlgrep ≠ Bentlybro/jevgrep ≠ can1357/jegrep ≠ uehaj/jev-semgrep", "default threshold 0.8 still soft", "40-line windows cannot prove whole function", "token-native sequential start/end Choice", "Gemini/Haiku stubs not configured yet", "handful of hand-written examples, not a benchmark", "Jev judged exactly what it was given", "laguagu/jev-skills ≠ laguagu/jev-evidence-lab ≠ Pleo2/awesome-jev-agent-skills", "contract_passed is not a claim of guaranteed factual truth", "Wilson lower bound 0.85 floor", "fixture mode no savings claim", "SemIf 2207★ (+21 vs §110 2186)", "jevlike 1043★ (+5 vs 1038)", "TypeAR 15★ (+1 vs 14)", "AnotiaWang 97★ (+1 vs 96)", "yibie/awesome-jev 506★ (+16 vs 490)", "Laya likes 822 (was 802)", "tracker likes 64 flat, lastModified UNCHANGED", "do not reopen or amend PR #23/#24/#25/#26/#27/#28", "Heman10x-NGU/Verdict-open-jev ≠ Heman10x-NGU/openJev-verdict-2.0", "TF-IDF + LogReg ECE 0.0207 vs Jev 0.1440", "Verdict-open-jev 48.07% vs Jev 90.80%", "abstention combined recall 10.00%", "p50 35.58 ms", "K=25 (maximum capacity) 72.00%", "0.85 coverage 84.60% selective risk 1.18%", "26.1× faster than standard Qwen JSON generation", "Jevify 90.0% / 167 ms CUDA graphs disabled", "Finding 1: Brier on stated confidence alone is a trap", "grpo_rlcr 0.78 / ECE 0.084", "reliability 0.007 but resolution 0.000", "27 900 schema-driven decisions", "13 600 / 13 600 questions", "candidate mass min 0.99999624", "22 configs · 166,054 rows · 4 calibration-gold", "sha a39eba3f", "Student B MAE 0.148 / Pearson 0.836 / 86.0%", "pngwn/open-jev-laya-bench README 404", "sha 9f69c742 likes 2", "HDFS 0.9933 (745/750) / retain 0.0084", "BGL ERROR/FATAL protection 1.0000", "2,479 / 2,500 HDFS uncertain", "cache hit 0.9648 (2412/2500)", "$0.153936 estimated", "E2 recomputes from saved probabilities", "Space sha eda59e0a", "MASSIVE English 0.783 / Khmer 0.033 / Hindi 0.133", "40–48 rows too small to ship T", "T never changes argmax", "siren2345/jev-single-decode-transformers ≠ siren2345/jev-single-decode", "Split Transformers experiment from llama.cpp runtime", "tanayvasishtha/jev-lab ≠ dairui1/jev-lab ≠ BrendanH18/jev-lab ≠ yibie/laya-jev-lab", "Four experiments stress-testing TypeSafe's Jev: calibration, bundle bias, label bias, and ensembling", "second pass must be $0.00 from cache", "The pages never call Jev", "Gemma 4 31B 77.0% / Jev 1.13.0 61.4% / Laya 322M 0.0%", "restriction state 95.0% against 84.4%", "None of the systems are particularly good at knowing when to stop and ask", "They skip the question and call a tool directly", "100% schema pass", "six-field joint 48.8% vs 72.8%", "ywchiu/jev_benchmark ≠ Running-Dolphins/jev-bench ≠ Praveenrajus/jev-bench", "ACT / REVIEW / FALLBACK", "A provider failure, timeout, malformed output, or missing answer is **not** a policy outcome", "confidence is descriptive provider output, not a substitute for probability", "Quality denominators include only valid scored answers", "an exact halfway tie chooses the lower level", "aiwithenoch/Jev-Skill ≠ simplosophy/jev-skill ≠ laguagu/jev-skills", "The local path does not claim to turn a smaller checkpoint into Jev", "Low support becomes decision: \"review\"", "MIT-0 SPDX NOASSERTION", "current-llm", "结构兼容，不是 Jev 模型能力", "altryne/jevify ≠ Mintzs/jevify ≠ gulagala001/jevify ≠ uspraveen/Jevify", "Find where Jev belongs. Design the questions. Measure the difference", "TypeAR-AI/TypeAR 301 → TypeLLM/TypeLLM", "TypeLLM/TypeLLM 16★", "SemIf 2241★ (+34 vs §111 2207)", "jevlike 1051★ (+8 vs 1043)", "AnotiaWang 98★ (+1 vs 97)", "yibie/awesome-jev 525★ (+19 vs 506)", "Laya likes 864 (was 822)", "tracker likes 67 (+3 vs 64)", "lastModified UNCHANGED `2026-09-20T04:29:16.000Z`", "do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#32", "hysteresis enter/exit / replay policy without inference", "calibration does not compose / hop-ECE permutation-invariant", "equal-width vs quantile ECE / ranking ≠ calibration", "Qwen2.5 ≠ Archer / Qwen 3.8 sparring ≠ Archer / Qwen/Qwen3.8-27B ≠ Archer", "Deferred Crispification / TCE / AMS", "g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev", "pd.cut equal-width vs jeval quantile", "A hunch is a probability with a policy attached", "soundness theater / measurement theater / hourly 0843", , "Jev Capability Resolver / NiazMorshed2007/jcr", "one tool nested capability tree / returns context / does not execute", "skills vs capabilities / workflow+judgment vs operations", "format independent of Jev / proposed open standard", "JCR_BAND_RATIO 0.6 is application policy / soft scores ≠ hard gates", "routing ≠ permission / docs ≠ authority to run", "sol-vs-opus5-20 lookup+explain / n=1 / Not Harbor task-execution", "wall-time mixed / Sol slower with JCR in 19/20", "NiazMorshed2007/jcr ≠ skill-broker ≠ skillranker ≠ jev-sift ≠ jev-lens ≠ jevusher ≠ jev_select_capability", "do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34", "notes.md §116", "copy the SemIf/MLX installer?", "quote 5.21× as beating Jev?", "treat 0.845 as a TypeSafe replica?", "collapse SemIf into kw2828/zhihz/semif-rs/semif-serve", "softmax over options as a Noul", "llm prompt to jev primitives", "conversion assistant not equivalent behavior", "heuristic conversion ≠ calibrated Noul", "alexwestco/llm-to-jev ≠ altryne/jevify", "user-provided 0940 / notes.md §118", "judge ≠ actuator", "candidate_mass", "softmax over A–H ≠ Noul", "hourly 0947 / notes.md §119", "ggmlc GGUF is not llama.cpp", "serving substrate ≠ calibrated replica", "Qwen3.5-9B ≠ Archer", "planner writes JEV selects", "hourly 1049 / notes.md §120", "open recreation ≠ calibrated replica", "semantic lint is a sensor not a proof", "cutoff 0.8 still soft", "paired bootstrap CIs *theirs*", "Same accuracy, 35x faster *theirs*", "hourly 1143 / notes.md §121", "revisit HIGH / since-last-look", "catalogued repo changed", "star-noise vs material change", "densify prior notes without inventing equivalence", "decide is not generate", "tryDecide returns typed calibrated judgments not a token stream", "GLiNER/GLiClass ports are class members not Jev replicas", "93.5% *theirs* not Harbor", "74.9 *theirs* not Harbor", "8.7x *theirs* not Harbor", "Option-Marker joint attention", "openjev:0.2.1", "thinking=True/False per-field budget", "PLAN_Qwen35", "hyperspaceai/jevcache ≠ kushals256/jevcache", "wire-compat ≠ logit-equiv", "SHA move is not a replica", "hourly 1248 / notes.md §123", "typesafe-sdk 0.7 Pydantic response models", "msgspec dropped", "The server's output is unchanged and was never wrong", "SchemaError is 400 plain-string detail not 422 list", "Pydantic response models ≠ logit-equiv", "msgspec dropped is not a replica", "Error contract is not a Noul", "coverage-at-error-budget *theirs* not Harbor", "PLAN_Qwen35 still proposal for review", "GLiNER locate ports are class members not Jev replicas", "Locate ≠ decide", "~160 ms *theirs* not Harbor", "0.971 F1 *theirs* not Harbor", "hf:fr0stbit3/laya-gguf serving substrate ≠ calibrated replica", "jkcdarunday/SystemOne-Next ≠ TypeSafe System One", "hourly 1340 / notes.md §124", "vLLM NVIDIA + MLX Apple Silicon", "Codiv hosted free endpoint", "dual /v1/systemone + /v1/chat/completions", "chat 501 on MLX", "dual serving is not generate", "Hosted Codiv ≠ TypeSafe", "hr98w/jev-visual 167★ Apple Silicon visual candidate scoring", "37.30s → 2.40s at 64 decisions *theirs*", "Breakout 9 bricks 6 returns 2 lives *theirs*", "candidate probabilities are relative not correctness", "jkudish/jev-mcp 156★ ten MCP tools", "recommendation is advisory", "the server never blocks on its own", "TypeSafe CLERC 5% to 18% *theirs*", "jkudish/jev-mcp ≠ burnigtm/jev-mcp", "zhengxuyu/litjev off-the-shelf Qwen decision layer", "Probabilities are not calibrated by default", "Qwen/Qwen3.8-27B ≠ Archer", "zhengxuyu/litjev ≠ alexwestco/llm-to-jev", "Zefan-Cai/Open-Jev LoRA + scalar head", "2B 94.71% 9B 97.54% hard test *theirs*", "2B OOD 86.02% 9B OOD 91.97% *theirs*", "80,816 training rows", "27B still in progress", "LoRA ≠ RLCD replica", "Zefan-Cai/Open-Jev ≠ TheoLeeCJ/openjev ≠ razorback16/openjev", "cristianoliveira/jeq intelligence you can pipe", "pass-min 0.8 still soft", "JEQ does not own actions", "AndyInQtr/laya-coreai CoreML serving substrate ≠ calibrated replica", "AndyInQtr/laya-coreai ≠ mizorewww/laya-coreml", "hourly 1441 / notes.md §125", "TypeLLM/TypeLLM densify HEAD 6a48f9f1e623", "README densify 3k→12k B", "Batch 5.8x *theirs*", "Constrained AR ≠ calibrated Noul", "jaredpalmer/kev densify HEAD b339f446a0ef", "Kev-0.6B 4B 8B family", "4B new-source 0.790/0.806 *theirs*", "8B new-source 0.796/0.780 *theirs*", "Jev hosted 0.857 *theirs*", "Questions share the input text but cannot read each other", "No Jev outputs were used for training", "8.2% ≥0.9 on wrong *theirs*", "option order can change an answer", "Qwen3 ≠ Archer", "TheoOliveira/pi-jev 21★ fail-closed routing", "JEV_THRESHOLD 0.65 still soft", "harshwasan/jev-sentinel fail closed never auto-allows", "harshwasan/jev-sentinel ≠ leepokai/jev-guard", "jackbarunz/jev-tool-router ≠ esinocchi/jev-tool-router", "threshold 0.90 still soft", "76/81 vs 77/81 *theirs*", "0.419s vs 2.459s *theirs*", "$0.00486 vs $0.03673 *theirs*", "not a security boundary", "baronunread/leanest fail-open uncertainty means RUN", "classifier.dev default Jev/Laya pluggable", "openlayer-ai/jevals ≠ dayhaysoos/jevals", "estimates not Harbor", "classifier ≠ authorizer", "MrJev/awesome-jev 118 entries catalog ≠ endorsement", "MrJev/awesome-jev ≠ yibie/awesome-jev", "Koushik890/jev-firewall fail closed ask_below 0.7 still soft", "CompleteTech-LLC-AI-Research/jev-codex-approval experimental native not compiled", "confidence is not a measured probability", "rh-guard owns primary gates", "hf:rAVEUK/open-jev-deberta-v3-large encoder class member not Jev replica", "hf:p-yan/laya-quanto serving substrate ≠ calibrated replica", "hf:Gtrkrsk/laya serving substrate ≠ calibrated replica", "hourly 1542 / notes.md §126", "razorback16/openjev densify HEAD febf02e88989", "release 0.3.0", "re-pin vLLM PR #57250 restructured head", "MODEL_VERSION stays openjev-0.1", "uv.lock hygiene", "restructured vLLM head ≠ logit-equiv", "frostney/clean-code-review 7★ typed judgments not opinions", "documentation is read not judged", "morcoan/JMP Joint Model Participation", "Models participate. Real tools execute.", "Thresholds are policy not model", "Kelbie/hunch ≠ carldaws/hunch ≠ tpellet/hunch ≠ huncho", "Jev never generates prose JSX or code", "json-render is the only renderer", "game success ≠ calibrated Noul", "Shalimov04/open-jev ≠ razorback16/openjev", "MstyAI/laya-onnx empty repo", "hf:Praveenrajus/jev-bench HTTP 200 was 401", "hourly 1643 / notes.md §127", "TypeLLM/TypeLLM densify HEAD 702e6a287f3c", "truncated thinking then constrained decode", "0.8B thinking On 0/18 *theirs*", "forced closure 20/20 type-valid *theirs*", "jaredpalmer/kev densify live HEAD 8465c4c4c294", "Kev-0.8B completes family", "4B new-source 0.794/0.832 *theirs*", "9B new-source 0.812/0.837 *theirs*", "transfer-v9 Kev-9B 5% Jev 9% Kev-8B 26% *theirs*", "SemIf Kev-9B 0.917 Jev 0.965 *theirs*", "scienthoon 0.952/0.911 vs 0.897/0.914 *theirs*", "transformers >= 5.17", "Qwen3.5 ≠ Archer", "notque/vexjoy-agent 421★ /d routes /do fallback", "Facts go to code. Judgments go to Jev. Only facts can block.", "Jev never blocks", "jqueryscript/awesome-jev ≠ MrJev/awesome-jev ≠ yibie/awesome-jev", "five-lines threshold 0.80 still soft", "371ms $0.0000189 300-call *theirs*", "tpellet/jevify ≠ altryne/jevify", "seb4ez/jevguard-mcp ≠ seb4ez/jevguard", "resumocast/jev-mcp ≠ jkudish/jev-mcp", "Adrian-Ernesto/jevsort ≠ zzzzzec/jevsort", "MidasMulli/kev-ane 155/155 argmax *theirs*", "hourly 1746 / notes.md §128", "Fine-tuning on your own data", "--data JSONL", "--init_from warm-start LoRA/head PR #9", "from-scratch ≠ warm-start", "JSONL labels ≠ Harbor", "Kev-0.8B 4B 9B Qwen3.5 family", "0.33 vs 0.84 vs 0.83/0.88 *theirs*", "reconstruction ≠ replica", "assay-001 split verdict", "Promethe-us/awesome-jev ≠ MrJev/awesome-jev ≠ yibie/awesome-jev", "ThePFMind/jev-mcp ≠ jkudish/jev-mcp", "kyegomez/open-jev ≠ razorback16/openjev", "namenu/pi-jev-effort ≠ TheoOliveira/pi-jev", "samatv256/mini-Jev ≠ r-ms/mini-jev", "hourly 1843 / notes.md §129", "TypeSafe-compatible ≠ TypeSafe replica", "SystemOne.from_pretrained", "replica ≠ TypeSafe", "76.7% vs Jev 86.9% strict common subset *theirs*", "kotoba-lang/typed-decisions ≠ convaiinnovations/laya-typed-decisions", "DeBERTa-v3-large 0.855 / 42 ms *theirs*", "aisearchio 15-link census catalog ≠ endorsement", "user-provided 1936 / notes.md §130", "systems latency ≠ semantic equivalence", "hard acc ≠ calibrated Noul", "Open-Jev TREC pending", "GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*", "TREC-DL Jev/Luna/Astra completed", "customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*", "1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*", "prefix caching experimental/off by default", "not merged base models", "Open-Jev densify HEAD 4933ee84951f", "Astra TREC commit 1dd56990be7e", "densify §125 not a sibling first sighting", "Open-Jev densify / notes.md §125", "launch X thread https://x.com/Zefan_Cai/status/2101782158658695388", "2101786019607740436", "2101789698947793231", platform does not execute trades / heyjunpenn/awesome-jev 485 catalog ≠ endorsement / 62.69% vs 67.26% *theirs* not gold / 203.2s $0.84 vs 823.5s $1.50 *theirs* / one seed-0 trial *theirs* / Jev $0.018825 vs Astra $5.93 *theirs* / 10.59× *theirs* / 6 class flips / agreement ≠ accuracy / probabilities uncalibrated / Qwen3.8 ≠ Archer / Spanish −6.4 pp XNLI *theirs* / ECE 0.057→0.101 *theirs* / 72.2% vs 63.4% p_max≥0.9 coverage *theirs* / Convert LLM prompts to Jev prompts / SHA unchanged 234058ab372d / skip Zefan-Cai/Open-Jev densify open #53 / skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54 / AI-reviewed labels ≠ gold / one-trial robot ≠ Harbor / 10.59× systems ≠ ECE / desc rewrite ≠ SHA/behavior change / rule-table ≠ model / local_only ≠ Jev / hourly 1946 / notes.md §131, or "cascade sign-flip / calibration theater": read `references/faq.md`, then `references/mental-models.md`, then `references/mixed-architecture.md`, then `references/judgment-class.md` before any mapping. Proof, @@ -592,3 +592,6 @@ Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 README SHA d331b567e2c3; SystemOne.from_pretrained; Batched single-token choice inference; TypeSafe-compatible; cache_prefix=True; LICENSE absent; sgoedecke/system-one ≠ mithalouni/system-one-open ≠ KathanModh259/system-one ≠ babybear-labs/system-one; TypeSafe-compatible ≠ TypeSafe replica; mithalouni/system-one-open 18★ MIT HEAD 77f1f7cccf8a README SHA 535f33028a68 LICENSE SHA 2f6f2cf1064e; Gemma 4 E2B / Gemma 3 270M Modal; 76.7% vs Jev 86.9% strict common subset *theirs*; 97 ms H100 *theirs*; 74.8% held-out *theirs*; replica ≠ TypeSafe; HF upload pending; kotoba-lang/typed-decisions 1★ Apache-2.0 HEAD 10d7834d3b99 README SHA 4d6bbf4c4e44 LICENSE SHA 513bb5e3cb4c; ModernBERT / DeBERTa / LLaDA-MoE; DeBERTa-v3-large 0.855 / 42 ms *theirs*; ModernBERT-base 0.717 / 68 ms *theirs*; LLaDA-MoE 0.835 / 676 ms *theirs*; kotoba-lang/typed-decisions ≠ convaiinnovations/laya-typed-decisions; encoder class member not Jev replica; aisearchio 15-link census catalog ≠ endorsement; 12 already carded 3 gaps this fold; soft scores ≠ hard gates; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §130 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 + +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/agent-self-assessment.md b/.agents/skills/augustus/references/agent-self-assessment.md index 054b559..bedb139 100644 --- a/.agents/skills/augustus/references/agent-self-assessment.md +++ b/.agents/skills/augustus/references/agent-self-assessment.md @@ -975,3 +975,5 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/applied-mappings.md b/.agents/skills/augustus/references/applied-mappings.md index abf578a..07c2243 100644 --- a/.agents/skills/augustus/references/applied-mappings.md +++ b/.agents/skills/augustus/references/applied-mappings.md @@ -2554,3 +2554,5 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/composition-algebra.md b/.agents/skills/augustus/references/composition-algebra.md index 3b1ccb6..af79e1e 100644 --- a/.agents/skills/augustus/references/composition-algebra.md +++ b/.agents/skills/augustus/references/composition-algebra.md @@ -3112,3 +3112,75 @@ Soft Noul ≠ hard safety. User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 README SHA d331b567e2c3; SystemOne.from_pretrained; Batched single-token choice inference; TypeSafe-compatible; cache_prefix=True; LICENSE absent; sgoedecke/system-one ≠ mithalouni/system-one-open ≠ KathanModh259/system-one ≠ babybear-labs/system-one; TypeSafe-compatible ≠ TypeSafe replica; mithalouni/system-one-open 18★ MIT HEAD 77f1f7cccf8a README SHA 535f33028a68 LICENSE SHA 2f6f2cf1064e; Gemma 4 E2B / Gemma 3 270M Modal; 76.7% vs Jev 86.9% strict common subset *theirs*; 97 ms H100 *theirs*; 74.8% held-out *theirs*; replica ≠ TypeSafe; HF upload pending; kotoba-lang/typed-decisions 1★ Apache-2.0 HEAD 10d7834d3b99 README SHA 4d6bbf4c4e44 LICENSE SHA 513bb5e3cb4c; ModernBERT / DeBERTa / LLaDA-MoE; DeBERTa-v3-large 0.855 / 42 ms *theirs*; ModernBERT-base 0.717 / 68 ms *theirs*; LLaDA-MoE 0.835 / 676 ms *theirs*; kotoba-lang/typed-decisions ≠ convaiinnovations/laya-typed-decisions; encoder class member not Jev replica; aisearchio 15-link census catalog ≠ endorsement; 12 already carded 3 gaps this fold; soft scores ≠ hard gates; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §130 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +505. **X-sentiment does not execute PRIMARY** (brainstormity/Jev-X-Sentiment-Analysis): + HEAD 5c932f941a92 README SHA bf4134b44cda. 136★. + platform does not execute trades. Buy/Sell/Hold/Take Profit. + Full cards: `judgment-class.md`, `validation.md`. +506. **awesome catalogs namesake** (heyjunpenn/awesome-jev): + heyjunpenn/awesome-jev 485 catalog ≠ endorsement. + heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one. + Full cards: `faq.md`. +507. **jev-arena *theirs* not gold** (NanmiCoder/jev-arena): + 10k comments 62.69% vs 67.26% *theirs* not gold. + 203.2s $0.84 vs 823.5s $1.50 *theirs*. AI-reviewed labels ≠ gold. + Full cards: `validation.md`. +508. **robot-control one-trial** (openroboto-ai/jev-robot-control): + one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. + one-trial robot ≠ Harbor. Full cards: `validation.md`. +509. **Qwen3.8 JevLike 10.59× uncalibrated** (endman100/research-Qwen3.8-JevLike): + 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. + probabilities uncalibrated. Qwen3.8 ≠ Archer. 10.59× systems ≠ ECE. + Full cards: `validation.md`, `faq.md`. +510. **jev-acento Spanish** (marcosmartinez/jev-acento): + Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. + 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. + Full cards: `validation.md`. +511. **llm-to-jev desc densify §118** (alexwestco/llm-to-jev): + description rewrite Convert LLM prompts to Jev prompts. + SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. + desc rewrite ≠ SHA/behavior change. Full cards: `question-design.md`. +512. **skip Open-Jev #53** (Zefan-Cai/Open-Jev): + skip Zefan-Cai/Open-Jev densify open #53. + Full cards: `faq.md`. +513. **skip #54 three** (sgoedecke / mithalouni / kotoba): + skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54. + Full cards: `faq.md`. +514. **already-carded §49/§62/§61** (ikermoel / nrdz-labs / mallahyari): + ikermoel/open-alternative-jev already §49. + nrdz-labs/fast-jev-opencode already §62. + mallahyari/system-one-benchmark already §61. + Full cards: `faq.md`. +515. **skill-suggester / MCP does not execute** (win4r/jev-skill-suggester, PyModel/typesafe-mcp): + does not execute. local_only ≠ Jev. host still reasons/edits/executes. + Full cards: `mixed-architecture.md`. +516. **JevLoop rule-table ≠ model** (Xubqpanda/JevLoop): + rule-table ≠ model. 12:1 / 7.7% *theirs* not Harbor. + Full cards: `validation.md`. +517. **omawish/rizzo-flow local ≠ replica** (elberacasa/omawish, Rizzo-AI-Academy/rizzo-flow): + replica ≠ TypeSafe. serving substrate ≠ calibrated replica. + 33M / 60/66 / 0/124 / 35ms *theirs*. ~250ms Q8 *theirs*. + Full cards: `judgment-class.md`. +518. **remainder apps** (live-jev / SEO / hub / labs / traders / others): + catalog ≠ endorsement. does not execute. + Full cards: `applied-mappings.md`, `faq.md`. +519. **namesake remainder** (awesome-jev / jev-mcp / mini-jev / OpenJev / typesafe-go / system-one / jevvy / jev-lab / pi / jev-harness): + catalog ≠ endorsement. SHA move is not a replica. + Full cards: `faq.md`. +520. **skip Archer** (promised_not_landed): + Qwen3.8 ≠ Archer. Hub archerhume/4rcherhume HTTP 401. + Archer still promised_not_landed. Full cards: `faq.md`. + +Hourly 1946 items 505–520 (`notes.md` §131). Do **not** +re-fold §129 items 481–496 / §128 items 465–480. +Leave 497–504 unused for open #54. +Skip Archer rewrite. Skip Open-Jev #53. +does not execute; catalog ≠ endorsement; +AI-reviewed labels ≠ gold; one-trial robot ≠ Harbor; +10.59× systems ≠ ECE; agreement ≠ accuracy; +desc rewrite ≠ SHA/behavior change; +SHA move is not a replica. +do not reopen or amend PR #23–#52. +Soft Noul ≠ hard safety. + +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/faq.md b/.agents/skills/augustus/references/faq.md index ce389c7..08984d2 100644 --- a/.agents/skills/augustus/references/faq.md +++ b/.agents/skills/augustus/references/faq.md @@ -3790,3 +3790,5 @@ Do not reopen or amend PR #23–#52. User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 README SHA d331b567e2c3; SystemOne.from_pretrained; Batched single-token choice inference; TypeSafe-compatible; cache_prefix=True; LICENSE absent; sgoedecke/system-one ≠ mithalouni/system-one-open ≠ KathanModh259/system-one ≠ babybear-labs/system-one; TypeSafe-compatible ≠ TypeSafe replica; mithalouni/system-one-open 18★ MIT HEAD 77f1f7cccf8a README SHA 535f33028a68 LICENSE SHA 2f6f2cf1064e; Gemma 4 E2B / Gemma 3 270M Modal; 76.7% vs Jev 86.9% strict common subset *theirs*; 97 ms H100 *theirs*; 74.8% held-out *theirs*; replica ≠ TypeSafe; HF upload pending; kotoba-lang/typed-decisions 1★ Apache-2.0 HEAD 10d7834d3b99 README SHA 4d6bbf4c4e44 LICENSE SHA 513bb5e3cb4c; ModernBERT / DeBERTa / LLaDA-MoE; DeBERTa-v3-large 0.855 / 42 ms *theirs*; ModernBERT-base 0.717 / 68 ms *theirs*; LLaDA-MoE 0.835 / 676 ms *theirs*; kotoba-lang/typed-decisions ≠ convaiinnovations/laya-typed-decisions; encoder class member not Jev replica; aisearchio 15-link census catalog ≠ endorsement; 12 already carded 3 gaps this fold; soft scores ≠ hard gates; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §130 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/formal-methods.md b/.agents/skills/augustus/references/formal-methods.md index 2f748bd..efb34cc 100644 --- a/.agents/skills/augustus/references/formal-methods.md +++ b/.agents/skills/augustus/references/formal-methods.md @@ -1404,3 +1404,5 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/formal-semi-formal.md b/.agents/skills/augustus/references/formal-semi-formal.md index 706a934..085a6ad 100644 --- a/.agents/skills/augustus/references/formal-semi-formal.md +++ b/.agents/skills/augustus/references/formal-semi-formal.md @@ -117,3 +117,5 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/judgment-class.md b/.agents/skills/augustus/references/judgment-class.md index eba2613..8a17db4 100644 --- a/.agents/skills/augustus/references/judgment-class.md +++ b/.agents/skills/augustus/references/judgment-class.md @@ -1519,3 +1519,5 @@ Open LM logit-trick (sgoedecke/system-one) is TypeSafe-compatible ≠ TypeSafe r User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 README SHA d331b567e2c3; SystemOne.from_pretrained; Batched single-token choice inference; TypeSafe-compatible; cache_prefix=True; LICENSE absent; sgoedecke/system-one ≠ mithalouni/system-one-open ≠ KathanModh259/system-one ≠ babybear-labs/system-one; TypeSafe-compatible ≠ TypeSafe replica; mithalouni/system-one-open 18★ MIT HEAD 77f1f7cccf8a README SHA 535f33028a68 LICENSE SHA 2f6f2cf1064e; Gemma 4 E2B / Gemma 3 270M Modal; 76.7% vs Jev 86.9% strict common subset *theirs*; 97 ms H100 *theirs*; 74.8% held-out *theirs*; replica ≠ TypeSafe; HF upload pending; kotoba-lang/typed-decisions 1★ Apache-2.0 HEAD 10d7834d3b99 README SHA 4d6bbf4c4e44 LICENSE SHA 513bb5e3cb4c; ModernBERT / DeBERTa / LLaDA-MoE; DeBERTa-v3-large 0.855 / 42 ms *theirs*; ModernBERT-base 0.717 / 68 ms *theirs*; LLaDA-MoE 0.835 / 676 ms *theirs*; kotoba-lang/typed-decisions ≠ convaiinnovations/laya-typed-decisions; encoder class member not Jev replica; aisearchio 15-link census catalog ≠ endorsement; 12 already carded 3 gaps this fold; soft scores ≠ hard gates; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §130 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/mappings.md b/.agents/skills/augustus/references/mappings.md index 651ad8a..fa412b3 100644 --- a/.agents/skills/augustus/references/mappings.md +++ b/.agents/skills/augustus/references/mappings.md @@ -2509,3 +2509,5 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/mental-models.md b/.agents/skills/augustus/references/mental-models.md index c734edb..8943336 100644 --- a/.agents/skills/augustus/references/mental-models.md +++ b/.agents/skills/augustus/references/mental-models.md @@ -3034,6 +3034,17 @@ Facts go to code. Judgments go to Jev. Only facts can block. Jev never blocks. five-lines threshold 0.80 still soft. 155/155 argmax *theirs*. serving substrate ≠ calibrated replica. SHA move is not a replica. Do not copy keys. +## Apply 1946 (`notes.md` §131) + +platform does not execute trades. catalog ≠ endorsement. +AI-reviewed labels ≠ gold. one-trial robot ≠ Harbor. +10.59× systems ≠ ECE. agreement ≠ accuracy. Qwen3.8 ≠ Archer. +Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. +72.2% vs 63.4% p_max≥0.9 coverage *theirs*. +desc rewrite ≠ SHA/behavior change. heuristic conversion ≠ calibrated Noul. +rule-table ≠ model. local_only ≠ Jev. replica ≠ TypeSafe. +SHA move is not a replica. Do not copy keys. + ## Apply 1843 (`notes.md` §129) from-scratch ≠ warm-start. JSONL labels ≠ Harbor. @@ -3152,3 +3163,5 @@ SHA move is not a replica. Do not reopen or amend PR #23–#52. User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 README SHA d331b567e2c3; SystemOne.from_pretrained; Batched single-token choice inference; TypeSafe-compatible; cache_prefix=True; LICENSE absent; sgoedecke/system-one ≠ mithalouni/system-one-open ≠ KathanModh259/system-one ≠ babybear-labs/system-one; TypeSafe-compatible ≠ TypeSafe replica; mithalouni/system-one-open 18★ MIT HEAD 77f1f7cccf8a README SHA 535f33028a68 LICENSE SHA 2f6f2cf1064e; Gemma 4 E2B / Gemma 3 270M Modal; 76.7% vs Jev 86.9% strict common subset *theirs*; 97 ms H100 *theirs*; 74.8% held-out *theirs*; replica ≠ TypeSafe; HF upload pending; kotoba-lang/typed-decisions 1★ Apache-2.0 HEAD 10d7834d3b99 README SHA 4d6bbf4c4e44 LICENSE SHA 513bb5e3cb4c; ModernBERT / DeBERTa / LLaDA-MoE; DeBERTa-v3-large 0.855 / 42 ms *theirs*; ModernBERT-base 0.717 / 68 ms *theirs*; LLaDA-MoE 0.835 / 676 ms *theirs*; kotoba-lang/typed-decisions ≠ convaiinnovations/laya-typed-decisions; encoder class member not Jev replica; aisearchio 15-link census catalog ≠ endorsement; 12 already carded 3 gaps this fold; soft scores ≠ hard gates; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §130 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/methods-catalog.md b/.agents/skills/augustus/references/methods-catalog.md index a857d28..dc122ba 100644 --- a/.agents/skills/augustus/references/methods-catalog.md +++ b/.agents/skills/augustus/references/methods-catalog.md @@ -311,3 +311,5 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/mixed-architecture.md b/.agents/skills/augustus/references/mixed-architecture.md index 0021707..96f789c 100644 --- a/.agents/skills/augustus/references/mixed-architecture.md +++ b/.agents/skills/augustus/references/mixed-architecture.md @@ -1492,3 +1492,5 @@ Batched single-token Choice after one prefill is still not hosted Jev. TypeSafe- User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 README SHA d331b567e2c3; SystemOne.from_pretrained; Batched single-token choice inference; TypeSafe-compatible; cache_prefix=True; LICENSE absent; sgoedecke/system-one ≠ mithalouni/system-one-open ≠ KathanModh259/system-one ≠ babybear-labs/system-one; TypeSafe-compatible ≠ TypeSafe replica; mithalouni/system-one-open 18★ MIT HEAD 77f1f7cccf8a README SHA 535f33028a68 LICENSE SHA 2f6f2cf1064e; Gemma 4 E2B / Gemma 3 270M Modal; 76.7% vs Jev 86.9% strict common subset *theirs*; 97 ms H100 *theirs*; 74.8% held-out *theirs*; replica ≠ TypeSafe; HF upload pending; kotoba-lang/typed-decisions 1★ Apache-2.0 HEAD 10d7834d3b99 README SHA 4d6bbf4c4e44 LICENSE SHA 513bb5e3cb4c; ModernBERT / DeBERTa / LLaDA-MoE; DeBERTa-v3-large 0.855 / 42 ms *theirs*; ModernBERT-base 0.717 / 68 ms *theirs*; LLaDA-MoE 0.835 / 676 ms *theirs*; kotoba-lang/typed-decisions ≠ convaiinnovations/laya-typed-decisions; encoder class member not Jev replica; aisearchio 15-link census catalog ≠ endorsement; 12 already carded 3 gaps this fold; soft scores ≠ hard gates; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §130 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/question-design.md b/.agents/skills/augustus/references/question-design.md index 85e10c7..9855e00 100644 --- a/.agents/skills/augustus/references/question-design.md +++ b/.agents/skills/augustus/references/question-design.md @@ -458,3 +458,5 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/toolbox-mapping.md b/.agents/skills/augustus/references/toolbox-mapping.md index aa3a08c..dc3d17b 100644 --- a/.agents/skills/augustus/references/toolbox-mapping.md +++ b/.agents/skills/augustus/references/toolbox-mapping.md @@ -387,3 +387,5 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/references/validation.md b/.agents/skills/augustus/references/validation.md index 3924b92..b2667a1 100644 --- a/.agents/skills/augustus/references/validation.md +++ b/.agents/skills/augustus/references/validation.md @@ -1194,3 +1194,5 @@ Hourly 1843 uniqueness lock: jaredpalmer/kev densify HEAD bd058057ad0a README SH User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 README SHA d331b567e2c3; SystemOne.from_pretrained; Batched single-token choice inference; TypeSafe-compatible; cache_prefix=True; LICENSE absent; sgoedecke/system-one ≠ mithalouni/system-one-open ≠ KathanModh259/system-one ≠ babybear-labs/system-one; TypeSafe-compatible ≠ TypeSafe replica; mithalouni/system-one-open 18★ MIT HEAD 77f1f7cccf8a README SHA 535f33028a68 LICENSE SHA 2f6f2cf1064e; Gemma 4 E2B / Gemma 3 270M Modal; 76.7% vs Jev 86.9% strict common subset *theirs*; 97 ms H100 *theirs*; 74.8% held-out *theirs*; replica ≠ TypeSafe; HF upload pending; kotoba-lang/typed-decisions 1★ Apache-2.0 HEAD 10d7834d3b99 README SHA 4d6bbf4c4e44 LICENSE SHA 513bb5e3cb4c; ModernBERT / DeBERTa / LLaDA-MoE; DeBERTa-v3-large 0.855 / 42 ms *theirs*; ModernBERT-base 0.717 / 68 ms *theirs*; LLaDA-MoE 0.835 / 676 ms *theirs*; kotoba-lang/typed-decisions ≠ convaiinnovations/laya-typed-decisions; encoder class member not Jev replica; aisearchio 15-link census catalog ≠ endorsement; 12 already carded 3 gaps this fold; soft scores ≠ hard gates; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §130 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/.agents/skills/augustus/scripts/evaluate_decisions.py b/.agents/skills/augustus/scripts/evaluate_decisions.py index 5e38689..32cae06 100755 --- a/.agents/skills/augustus/scripts/evaluate_decisions.py +++ b/.agents/skills/augustus/scripts/evaluate_decisions.py @@ -574,6 +574,46 @@ def latency_is_not_task_quality(quality_claimed=False): return quality_claimed is False +def platform_does_not_execute(executes=False): + """Dashboard Choice is not a fill.""" + return executes is False + + +def ai_reviewed_is_not_gold(ai_reviewed=True, gold=False): + """AI-reviewed labels ≠ gold.""" + return ai_reviewed is True and gold is False + + +def one_trial_is_not_harbor(n_trials, harbor=False): + """one seed-0 trial ≠ Harbor.""" + if n_trials != 1: + raise ValueError("unexpected n") + return harbor is False + + +def systems_speedup_is_not_ece(kind, ece_claimed=False): + """10.59× systems comparison ≠ ECE.""" + if kind != "systems_timing": + raise ValueError("unexpected kind") + return ece_claimed is False + + +def agreement_is_not_accuracy(agree, labeled=False): + """agreement ≠ accuracy when there are no labels.""" + return agree is True and labeled is False + + +def desc_rewrite_is_not_sha_change(desc_changed, sha_changed=False): + """desc rewrite ≠ SHA/behavior change.""" + return desc_changed is True and sha_changed is False + + +def rule_table_is_not_model(source, model_claimed=False): + """rule-table ≠ model.""" + if source != "rule_table": + raise ValueError("unexpected source") + return model_claimed is False + def hop_ece_permutation_invariant(rows, bins=10, key="p"): """Shuffle order; equal-width ECE must not move. @@ -820,6 +860,27 @@ def self_test(): assert latency_is_not_task_quality(False) assert not latency_is_not_task_quality(True) + # 1946: does not execute / AI-reviewed ≠ gold / one-trial ≠ Harbor / + # 10.59× ≠ ECE / agreement ≠ accuracy / desc rewrite ≠ SHA. + assert platform_does_not_execute(False) + assert not platform_does_not_execute(True) + assert ai_reviewed_is_not_gold(True, False) + assert not ai_reviewed_is_not_gold(True, True) + assert one_trial_is_not_harbor(1, False) + assert not one_trial_is_not_harbor(1, True) + assert systems_speedup_is_not_ece("systems_timing", False) + assert not systems_speedup_is_not_ece("systems_timing", True) + assert agreement_is_not_accuracy(True, False) + assert not agreement_is_not_accuracy(True, True) + assert desc_rewrite_is_not_sha_change(True, False) + assert not desc_rewrite_is_not_sha_change(True, True) + assert rule_table_is_not_model("rule_table", False) + assert not rule_table_is_not_model("rule_table", True) + assert theirs_bench_is_not_harbor(10000, "jev-arena-10k") + assert theirs_bench_is_not_harbor(1, "robot-seed-0") + assert theirs_bench_is_not_harbor(10, "qwen38-jevlike-10.59x") + assert theirs_bench_is_not_harbor(3200, "jev-acento-paired") + print("self-test ok") diff --git a/.agents/skills/augustus/scripts/uniqueness_gate.py b/.agents/skills/augustus/scripts/uniqueness_gate.py index 4331984..faa62c3 100644 --- a/.agents/skills/augustus/scripts/uniqueness_gate.py +++ b/.agents/skills/augustus/scripts/uniqueness_gate.py @@ -2,7 +2,7 @@ """Uniqueness gate for merged 0843 (§114), merged 0915 NanoJev (§115), merged 0920 jcr (§116), merged 0922 SemIf (§117), merged 0940 llm-to-jev (§118), hourly 0947 HIGH (§119), hourly 1049 HIGH (§120), -hourly 1143 HIGH (§121), hourly 1248 HIGH (§123), hourly 1340 HIGH (§124), hourly 1441 HIGH (§125), hourly 1542 HIGH (§126), hourly 1643 HIGH (§127), hourly 1746 HIGH (§128), hourly 1843 HIGH (§129), user-provided 1936 HIGH (§130), and Open-Jev densify (§125). +hourly 1143 HIGH (§121), hourly 1248 HIGH (§123), hourly 1340 HIGH (§124), hourly 1441 HIGH (§125), hourly 1542 HIGH (§126), hourly 1643 HIGH (§127), hourly 1746 HIGH (§128), hourly 1843 HIGH (§129), user-provided 1936 HIGH (§130), Open-Jev densify (§125), and hourly 1946 HIGH (§131). Each lock must appear as one consecutive substring in every listed overlay. Fragments scattered across files do not count. @@ -11,10 +11,11 @@ substring in the skill + research files (not a 21-overlay dump wall). Hourly must treat revisit HIGH like novel HIGH. Star-noise is not a fold. -Also: YAML-parse SKILL.md frontmatter; notes.md owns §114–§130; -composition items 289–316, 322–329, 330–336, 337–352, 353–368, 369–384, 385–400, 401–416, 417–432, 433–448, 449–464, 465–480, 481–496, and 497–504 exist; -findings batches #97–#112 exist. Items 317–321 stay unused. -The 1843 archive run_digest must claim §129 / 481–496 / #111 +Also: YAML-parse SKILL.md frontmatter; notes.md owns §114–§131; +composition items 289–316, 322–329, 330–336, 337–352, 353–368, 369–384, 385–400, 401–416, 417–432, 433–448, 449–464, 465–480, 481–496, 497–504, and 505–520 exist; +findings batches #97–#113 exist. Items 317–321 stay unused. +The 1843 archive run_digest must claim §129 / 481–496 / #111. +The 1946 archive run_digest must claim §131 / 505–520 / #113 (not the 1746 IDs §128 / 465–480 / #110). CHANGELOG.md must not hold uniqueness dump walls (dumps live in changelog-hourly.md). README.md must not hold the 0743 dump wall. @@ -173,6 +174,10 @@ 'User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125' ) +UNIQ_1946 = ( + 'Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131' +) + REVISIT_LOCK = ( "Revisit / since-last-look lock: catalogued repos are not done; " "store fingerprints default_sha, pushed_at, description_hash, release_tag; " @@ -270,6 +275,8 @@ def main() -> int: failed.append(f"1936 lock missing as one substring: {rel}") if UNIQ_OPENJEV not in body: failed.append(f"openjev densify lock missing as one substring: {rel}") + if UNIQ_1946 not in body: + failed.append(f"1946 lock missing as one substring: {rel}") for rel in REVISIT_OVERLAYS: path = ROOT / rel if not path.is_file(): @@ -313,10 +320,12 @@ def main() -> int: failed.append("notes.md missing §129 heading") if "## 130. User-provided HIGH" not in notes: failed.append("notes.md missing §130 heading") + if "## 131. Hourly 1946 HIGH" not in notes: + failed.append("notes.md missing §131 heading") algebra = (ROOT / ".agents/skills/augustus/references/composition-algebra.md").read_text( encoding="utf-8" ) - for n in list(range(289, 317)) + list(range(322, 330)) + list(range(330, 337)) + list(range(337, 353)) + list(range(353, 369)) + list(range(369, 385)) + list(range(385, 401)) + list(range(401, 417)) + list(range(417, 433)) + list(range(433, 449)) + list(range(449, 465)) + list(range(465, 481)) + list(range(481, 497)) + list(range(497, 505)): + for n in list(range(289, 317)) + list(range(322, 330)) + list(range(330, 337)) + list(range(337, 353)) + list(range(353, 369)) + list(range(369, 385)) + list(range(385, 401)) + list(range(401, 417)) + list(range(417, 433)) + list(range(433, 449)) + list(range(449, 465)) + list(range(465, 481)) + list(range(481, 497)) + list(range(497, 505)) + list(range(505, 521)): needle = f"{n}. **" if needle not in algebra: failed.append(f"composition-algebra missing item {n}") @@ -342,10 +351,32 @@ def main() -> int: "## Batch #110", "## Batch #111", "## Batch #112", + "## Batch #113", ): if batch not in findings: failed.append(f"findings.md missing {batch}") digest_path = ROOT / "research/archive/hourly/2026-09-21T00/run_digest.json" + digest_path_1946 = ROOT / "research/archive/hourly/2026-09-21T01/run_digest.json" + if not digest_path_1946.is_file(): + failed.append("missing 1946 run_digest.json") + else: + digest1946 = json.loads(digest_path_1946.read_text(encoding="utf-8")) + if digest1946.get("label") != "1946": + failed.append(f"1946 run_digest label {digest1946.get('label')!r} != '1946'") + if digest1946.get("notes_section") != "131": + failed.append( + f"1946 run_digest notes_section {digest1946.get('notes_section')!r} != '131'" + ) + if digest1946.get("composition") != "505-520": + failed.append( + f"1946 run_digest composition {digest1946.get('composition')!r} != '505-520'" + ) + if digest1946.get("findings_batch") != 113: + failed.append( + f"1946 run_digest findings_batch {digest1946.get('findings_batch')!r} != 113" + ) + if digest1946.get("invented_signal") is not False: + failed.append("1946 run_digest invented_signal is not false") if not digest_path.is_file(): failed.append("missing 1843 run_digest.json") else: @@ -625,6 +656,31 @@ def main() -> int: "launch X thread https://x.com/Zefan_Cai/status/2101782158658695388", "2101786019607740436", "2101789698947793231", + 'platform does not execute trades', + 'heyjunpenn/awesome-jev 485 catalog ≠ endorsement', + '62.69% vs 67.26% *theirs* not gold', + '203.2s $0.84 vs 823.5s $1.50 *theirs*', + 'one seed-0 trial *theirs*', + 'Jev $0.018825 vs Astra $5.93 *theirs*', + '10.59× *theirs*', + '6 class flips', + 'agreement ≠ accuracy', + 'probabilities uncalibrated', + 'Qwen3.8 ≠ Archer', + 'Spanish −6.4 pp XNLI *theirs*', + 'ECE 0.057→0.101 *theirs*', + '72.2% vs 63.4% p_max≥0.9 coverage *theirs*', + 'Convert LLM prompts to Jev prompts', + 'SHA unchanged 234058ab372d', + 'skip Zefan-Cai/Open-Jev densify open #53', + 'skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54', + 'hourly 1946 / notes.md §131', + 'AI-reviewed labels ≠ gold', + 'one-trial robot ≠ Harbor', + '10.59× systems ≠ ECE', + 'desc rewrite ≠ SHA/behavior change', + 'rule-table ≠ model', + 'local_only ≠ Jev', ): if frag not in haystack: failed.append(f"SKILL.md missing fragment {frag!r}") @@ -845,6 +901,31 @@ def main() -> int: "launch X thread https://x.com/Zefan_Cai/status/2101782158658695388", "2101786019607740436", "2101789698947793231", + 'platform does not execute trades', + 'heyjunpenn/awesome-jev 485 catalog ≠ endorsement', + '62.69% vs 67.26% *theirs* not gold', + '203.2s $0.84 vs 823.5s $1.50 *theirs*', + 'one seed-0 trial *theirs*', + 'Jev $0.018825 vs Astra $5.93 *theirs*', + '10.59× *theirs*', + '6 class flips', + 'agreement ≠ accuracy', + 'probabilities uncalibrated', + 'Qwen3.8 ≠ Archer', + 'Spanish −6.4 pp XNLI *theirs*', + 'ECE 0.057→0.101 *theirs*', + '72.2% vs 63.4% p_max≥0.9 coverage *theirs*', + 'Convert LLM prompts to Jev prompts', + 'SHA unchanged 234058ab372d', + 'skip Zefan-Cai/Open-Jev densify open #53', + 'skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54', + 'hourly 1946 / notes.md §131', + 'AI-reviewed labels ≠ gold', + 'one-trial robot ≠ Harbor', + '10.59× systems ≠ ECE', + 'desc rewrite ≠ SHA/behavior change', + 'rule-table ≠ model', + 'local_only ≠ Jev', ): if frag not in proto_line: failed.append(f"SKILL.md protocol missing {frag!r}") @@ -867,6 +948,7 @@ def main() -> int: ("1843", UNIQ_1843), ("1936", UNIQ_1936), ("openjev_densify", UNIQ_OPENJEV), + ("1946", UNIQ_1946), ): if lock in changelog: failed.append( @@ -945,6 +1027,7 @@ def main() -> int: f"1843 chars={len(UNIQ_1843)} " f"1936 chars={len(UNIQ_1936)} " f"openjev_densify chars={len(UNIQ_OPENJEV)} " + f"1946 chars={len(UNIQ_1946)} " f"revisit chars={len(REVISIT_LOCK)} " f"overlays={len(OVERLAYS)} " f"revisit_overlays={len(REVISIT_OVERLAYS)}" diff --git a/CHANGELOG.md b/CHANGELOG.md index 882f248..5afc649 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -16,6 +16,36 @@ folds: `research/notes.md`. ## [Unreleased] +Hourly 1946 HIGH (`research/notes.md` §131 / composition items +505–520 / findings batch #113). Does **not** bump the 0.5.0 pin. +Uniqueness dumps live in +[`research/changelog-hourly.md`](research/changelog-hourly.md). +Do not reopen or amend PR #23–#52. +Do not amend released 0.5.0 (#42). Merged #53 owns Open-Jev densify on §125. Merged #54 owns §130. Merged #52 owns §129. + +### Added + +- **Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute / + heyjunpenn 485 catalog ≠ endorsement / jev-arena 62.69% vs 67.26% + *theirs* not gold / one seed-0 robot trial *theirs* / + 10.59× systems ≠ ECE / Spanish −6.4 pp XNLI *theirs* / + llm-to-jev description rewrite SHA unchanged. + Evaluator: does not execute / AI-reviewed ≠ gold / one-trial ≠ Harbor / + 10.59× ≠ ECE / agreement ≠ accuracy / desc rewrite ≠ SHA. + uniqueness_gate.py now checks 0843 + 0915 + jcr + 0922 + 0940 + 0947 + + 1049 + 1143 + 1248 + 1340 + 1441 + 1542 + 1643 + 1746 + 1843 + 1936 + Open-Jev densify + 1946. + Composition items 505–520 / batch #113. + **HARD RULE:** do not reopen or amend PR #23–#52. Does **not** bump + 0.5.0. + +- **Recipe (class, not Jev-only).** Without Augustus: treat a dashboard + Choice as a fill, a catalog as a grant, AI-reviewed labels as gold, a + one-trial robot run as Harbor, 10.59× as ECE, or a description rewrite + as a SHA change. With Augustus: does not execute; catalog ≠ endorsement; + AI-reviewed labels ≠ gold; one-trial robot ≠ Harbor; 10.59× systems ≠ + ECE; desc rewrite ≠ SHA/behavior change. Same split for any + Choice/Score/Noul-style head, not only hosted Jev. + Open-Jev densify (`research/notes.md` §125). Does **not** bump the 0.5.0 pin. Uniqueness dumps live in [`research/changelog-hourly.md`](research/changelog-hourly.md). diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index 719e45c..997423a 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -45,7 +45,7 @@ re-opened as "new." Before folding: `.agents/skills/augustus/SKILL.md` - Do not re-fold an already-landed section as a new beat - Do not reopen or amend a merged fold PR (#23–#52) -- uniqueness_gate.py checks 0843 + 0915 + jcr + 0922 + 0940 + 0947 + 1049 + 1143 + 1248 + 1340 + 1441 + 1542 + 1643 + 1746 + 1843 + 1936 consecutive locks, plus the revisit / since-last-look protocol substring in the skill and research files. +- uniqueness_gate.py checks 0843 + 0915 + jcr + 0922 + 0940 + 0947 + 1049 + 1143 + 1248 + 1340 + 1441 + 1542 + 1643 + 1746 + 1843 + 1936 + openjev + 1946 consecutive locks, plus the revisit / since-last-look protocol substring in the skill and research files. - Hourly uniqueness dump: `research/changelog-hourly.md` (archive, not release notes) - Treat **revisit HIGH like novel HIGH**. Catalogued repos are not diff --git a/README.md b/README.md index ff4342e..364f1cb 100644 --- a/README.md +++ b/README.md @@ -185,3 +185,5 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/docs/_includes/recipes.html b/docs/_includes/recipes.html index 1ebe8f9..20e1ae4 100644 --- a/docs/_includes/recipes.html +++ b/docs/_includes/recipes.html @@ -335,6 +335,48 @@

Latency vs meaning

customer-service P50 85.03 vs 295.26 *theirs*. 1024/32 1015.90 vs 301.37 *theirs*. Open-Jev TREC pending.
+
+

Decision support

+

Dashboard vs fill

+
+
Problem
+
A Buy/Sell/Hold card treated as an executed order.
+
Without
+
Wire the Choice to the exchange. Skip the human.
+
With
+
platform does not execute trades. does not execute. Code owns the fill.
+
Measure
+
Count fills separately from cards. *theirs* not Harbor.
+
+
+
+

Independent assay

+

AI review vs gold / one trial vs Harbor

+
+
Problem
+
62.69% or one seed-0 robot run treated as a certificate.
+
Without
+
Ship AI labels as gold. Quote $0.018825 as Harbor.
+
With
+
AI-reviewed labels ≠ gold. one-trial robot ≠ Harbor. 10.59× systems ≠ ECE. agreement ≠ accuracy.
+
Measure
+
62.69% vs 67.26% *theirs* not gold. Spanish −6.4 pp XNLI *theirs*. Not Harbor.
+
+
+
+

On-ramp / catalog

+

Desc rewrite vs replica

+
+
Problem
+
A GitHub description rewrite treated as a new compiler, or a catalog as a grant.
+
Without
+
Mint a sibling card. Quote 485 as endorsement.
+
With
+
desc rewrite ≠ SHA/behavior change. catalog ≠ endorsement. heuristic conversion ≠ calibrated Noul.
+
Measure
+
SHA unchanged 234058ab372d. 485 is an index. *theirs* not Harbor.
+
+

Measurement recipe (hysteresis, equal-width vs quantile ECE, hop-ECE, diff --git a/docs/ecosystem.md b/docs/ecosystem.md index 64bebe1..6a2df9f 100644 --- a/docs/ecosystem.md +++ b/docs/ecosystem.md @@ -1176,3 +1176,5 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +**Hourly 1946 HIGH (`notes.md` §131).** X-sentiment does not execute trades. heyjunpenn/awesome-jev 485 catalog ≠ endorsement. jev-arena 62.69% vs 67.26% *theirs* not gold. 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. Spanish −6.4 pp XNLI *theirs*. ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. llm-to-jev description rewrite Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. heuristic conversion ≠ calibrated Noul. skip Zefan-Cai/Open-Jev densify open #53. skip #54 three. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/research/archive/findings.md b/research/archive/findings.md index 80f126f..101afaf 100644 --- a/research/archive/findings.md +++ b/research/archive/findings.md @@ -2,6 +2,36 @@ +## Batch #113 (2026-09-20 ~19:46 Boise / ~01:46 UTC) - hourly 1946 HIGH + +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 + +Note: `research/notes.md` §131. Docs + evaluator, fresh PR off latest +`main` (`6a7a557` / merged #54 aisearchio §130). Merged #54 owns §130 / +497–504 / #112. Merged #52 owns §129. Merged #51 owns §128. Open #53 owns +Open-Jev densify. This fold stays §131 / items 505–520 / batch #113. +**HARD RULE:** do not reopen or amend PR #23–#52. +Quote READMEs. Soft Noul ≠ hard safety. Augustus owns +placement. `invented_signal: false`. + +- **X-sentiment / catalogs PRIMARY.** HEAD 5c932f941a92. platform does not execute trades. + heyjunpenn/awesome-jev 485 catalog ≠ endorsement. +- **jev-arena / robot-control / JevLike.** 62.69% vs 67.26% *theirs* not gold. + 203.2s $0.84 vs 823.5s $1.50 *theirs*. one seed-0 trial *theirs*. + Jev $0.018825 vs Astra $5.93 *theirs*. 10.59× *theirs*. 6 class flips. + agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. +- **jev-acento / llm-to-jev densify.** Spanish −6.4 pp XNLI *theirs*. + ECE 0.057→0.101 *theirs*. 72.2% vs 63.4% p_max≥0.9 coverage *theirs*. + Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. 3★. +- **skips / namesakes.** skip Zefan-Cai/Open-Jev densify open #53. + skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54. + ikermoel already §49. nrdz-labs already §62. catalog ≠ endorsement. + Archer still promised_not_landed. + +Pulse: Archer still NOT landed. Hub archerhume/4rcherhume HTTP **401**. +Jev-X-Sentiment-Analysis **136★**. awesome-jev **32★**. jev-arena **31★**. +llm-to-jev **3★**. `invented_signal: false`. + ## Batch #112 (2026-09-20 ~19:36 Boise / 2026-09-21T01:36Z) - user-provided 1936 HIGH User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 README SHA d331b567e2c3; SystemOne.from_pretrained; Batched single-token choice inference; TypeSafe-compatible; cache_prefix=True; LICENSE absent; sgoedecke/system-one ≠ mithalouni/system-one-open ≠ KathanModh259/system-one ≠ babybear-labs/system-one; TypeSafe-compatible ≠ TypeSafe replica; mithalouni/system-one-open 18★ MIT HEAD 77f1f7cccf8a README SHA 535f33028a68 LICENSE SHA 2f6f2cf1064e; Gemma 4 E2B / Gemma 3 270M Modal; 76.7% vs Jev 86.9% strict common subset *theirs*; 97 ms H100 *theirs*; 74.8% held-out *theirs*; replica ≠ TypeSafe; HF upload pending; kotoba-lang/typed-decisions 1★ Apache-2.0 HEAD 10d7834d3b99 README SHA 4d6bbf4c4e44 LICENSE SHA 513bb5e3cb4c; ModernBERT / DeBERTa / LLaDA-MoE; DeBERTa-v3-large 0.855 / 42 ms *theirs*; ModernBERT-base 0.717 / 68 ms *theirs*; LLaDA-MoE 0.835 / 676 ms *theirs*; kotoba-lang/typed-decisions ≠ convaiinnovations/laya-typed-decisions; encoder class member not Jev replica; aisearchio 15-link census catalog ≠ endorsement; 12 already carded 3 gaps this fold; soft scores ≠ hard gates; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §130 @@ -34,7 +64,6 @@ Pulse: Archer still NOT landed. Hub archerhume/4rcherhume HTTP **401**. sgoedecke **20★**. mithalouni **18★**. kotoba-lang/typed-decisions **1★**. `invented_signal: false`. - ## Batch #111 (2026-09-20 ~18:43 Boise / ~00:43 UTC) - hourly 1843 HIGH Hourly 1843 uniqueness lock: jaredpalmer/kev densify HEAD bd058057ad0a README SHA 84b872488915; Fine-tuning on your own data; --data JSONL; --init_from warm-start LoRA/head PR #9; Kev-0.8B 4B 9B Qwen3.5 family; 4B new-source 0.794/0.832 *theirs*; 9B new-source 0.812/0.837 *theirs*; 0.33 vs 0.84 vs 0.83/0.88 *theirs*; from-scratch ≠ warm-start; JSONL labels ≠ Harbor; Kev-0.5B card Qwen3.5 family pointer; No Jev outputs were used for training; option order can change an answer; 8.2% ≥0.9 on wrong *theirs*; Kev-9B 7.5% ≥0.9 on wrong *theirs*; dabit3/jev-experiments densify 340★; simota/tenbin densify neighbor skill; Promethe-us/awesome-jev ≠ MrJev/awesome-jev ≠ yibie/awesome-jev; kyegomez/open-jev reconstruction ≠ replica; unofficial research implementation with random weights; kyegomez/open-jev ≠ razorback16/openjev ≠ TheoLeeCJ/openjev ≠ Zefan-Cai/Open-Jev ≠ Shalimov04/open-jev; jourdanlabs/assay-001 split verdict; CLINC150 ECE 0.0204 *theirs*; Banking77 ECE 0.0936 *theirs*; 8,576 responses zero type errors *theirs*; brnyxx/jev-ra 3-5x / ~300 ms *theirs*; 8.50× Wikipedia *theirs*; ThePFMind/jev-mcp ≠ jkudish/jev-mcp ≠ burnigtm/jev-mcp; namenu/pi-jev-effort ≠ TheoOliveira/pi-jev; samatv256/mini-Jev ≠ r-ms/mini-jev; comoc/jev-minesweeper ≠ EnesYilmazcode/JevMinesweeper; game success ≠ calibrated Noul; Nutlope/jev-fraud Kimi K3; jeffloo886/jev-notion; hf:akhilaaa3/openjev-v1-allmix-r512-merged ≠ hf:akhilaaa3/openjev-r512-handoff-demo; serving substrate ≠ calibrated replica; catalog ≠ endorsement; SHA move is not a replica; wire-compat ≠ logit-equiv; Qwen3.5 ≠ Archer; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51; notes.md §129 diff --git a/research/archive/hourly/2026-09-21T01/augustus_items.json b/research/archive/hourly/2026-09-21T01/augustus_items.json new file mode 100644 index 0000000..8d23d27 --- /dev/null +++ b/research/archive/hourly/2026-09-21T01/augustus_items.json @@ -0,0 +1,668 @@ +[ + { + "id": "brainstormity/Jev-X-Sentiment-Analysis", + "kind": "novel", + "description": "", + "stars": 135, + "why": null, + "html_url": "https://github.com/brainstormity/Jev-X-Sentiment-Analysis", + "bench_mark": null + }, + { + "id": "ikermoel/open-alternative-jev", + "kind": "novel", + "description": "Open-source alternative to TypeSafe's Jev: a System One style model layer that gives typed, calibrated decisions from any open-weights LLM in one forward pass (HF + vLLM), with honest benchmarks", + "stars": 36, + "why": null, + "html_url": "https://github.com/ikermoel/open-alternative-jev", + "bench_mark": null + }, + { + "id": "okinaaudio/live-jev", + "kind": "novel", + "description": "Control Ableton Live with one short sentence (Japanese / English). Summon with \u2318\u21e7Space, type or dictate, done.", + "stars": 36, + "why": null, + "html_url": "https://github.com/okinaaudio/live-jev", + "bench_mark": null + }, + { + "id": "heyjunpenn/awesome-jev", + "kind": "novel", + "description": "A verified, community-maintained catalog of 503 open-source projects built with Jev.", + "stars": 32, + "why": null, + "html_url": "https://github.com/heyjunpenn/awesome-jev", + "bench_mark": null + }, + { + "id": "NanmiCoder/jev-arena", + "kind": "novel", + "description": "Jev \u6a21\u578b\u4ecb\u7ecd\u4e0e\u5b9e\u6d4b\uff1a\u901a\u8fc7 Choice / Score / Noul \u5c06\u81ea\u7136\u8bed\u8a00\u8f6c\u4e3a\u5e26\u7c7b\u578b\u7684\u5224\u65ad\u4e0e\u6982\u7387\uff0c\u7528\u4e8e\u5206\u7c7b\u3001\u8bc4\u5206\u548c\u8def\u7531\uff1b\u652f\u6301\u4e0e DeepSeek \u7b49\u6a21\u578b\u5bf9\u6bd4\u8bc4\u8bba\u6253\u6807\u3001\u901f\u5ea6\u4e0e\u7ed3\u679c\uff0c\u542b CSV/Excel \u5bfc\u5165\u3001\u539f\u901f\u56de\u653e\u4e0e\u79bb\u7ebf\u62a5\u544a\u3002", + "stars": 31, + "why": null, + "html_url": "https://github.com/NanmiCoder/jev-arena", + "bench_mark": null + }, + { + "id": "yzfly/awesome-jev-zh", + "kind": "novel", + "description": "Jev / TypeSafe System One \u4e2d\u6587\u7cbe\u9009\u5217\u8868\uff1a\u5b98\u65b9\u8d44\u6599\u3001SDK\u3001\u7206\u6b3e\u5e94\u7528\u3001Agent \u5de5\u5177\u3001\u5f00\u6e90\u590d\u73b0\u4e0e\u72ec\u7acb\u8bc4\u6d4b\uff0c\u9644\u4e2d\u6587\u4e0a\u624b\u6307\u5357\uff0c\u6bcf\u65e5\u81ea\u52a8\u6536\u5f55 GitHub \u70ed\u95e8\u9879\u76ee\u3002", + "stars": 31, + "why": null, + "html_url": "https://github.com/yzfly/awesome-jev-zh", + "bench_mark": null + }, + { + "id": "openroboto-ai/jev-robot-control", + "kind": "novel", + "description": "", + "stars": 27, + "why": null, + "html_url": "https://github.com/openroboto-ai/jev-robot-control", + "bench_mark": null + }, + { + "id": "win4r/jev-skill-suggester", + "kind": "novel", + "description": "\u7528 TypeSafe Jev \u63a8\u8350\u5df2\u5b89\u88c5 Skill / Bounded installed-skill recommendations with TypeSafe Jev. Python CLI, Codex skill, bilingual docs and live examples.", + "stars": 27, + "why": null, + "html_url": "https://github.com/win4r/jev-skill-suggester", + "bench_mark": null + }, + { + "id": "PyModel/typesafe-mcp", + "kind": "novel", + "description": "", + "stars": 23, + "why": null, + "html_url": "https://github.com/PyModel/typesafe-mcp", + "bench_mark": null + }, + { + "id": "AkashPriyadarshii/jev-seo", + "kind": "novel", + "description": "100% free \u20b90 agent-first SEO & GEO CLI suite and MCP server in Rust replacing Semrush and OpenSEO via DuckDuckGo and TypeSafe Jev System One", + "stars": 21, + "why": null, + "html_url": "https://github.com/AkashPriyadarshii/jev-seo", + "bench_mark": null + }, + { + "id": "devtooligan/jevscan-evm", + "kind": "novel", + "description": "", + "stars": 21, + "why": null, + "html_url": "https://github.com/devtooligan/jevscan-evm", + "bench_mark": null + }, + { + "id": "mizzlelover/jev-hub", + "kind": "novel", + "description": "JEV HUB \u00b7 X \u4e0a\u5173\u4e8e TypeSafe AI\u300c\u7cfb\u7edf\u4e00\u6a21\u578b\u300dJev \u7684\u957f\u6587\u4e0e\u6f14\u793a\u89c6\u9891\u805a\u5408\uff08\u4fdd\u7559\u539f\u94fe\u4e0e\u4f5c\u8005\uff09\uff5c \u8c01\u662f\u4e13\u5bb6 \u51fa\u54c1", + "stars": 21, + "why": null, + "html_url": "https://github.com/mizzlelover/jev-hub", + "bench_mark": null + }, + { + "id": "sgoedecke/system-one", + "kind": "novel", + "description": "Batched single-token choice inference for open language models, compatible with TypeSafe", + "stars": 20, + "why": null, + "html_url": "https://github.com/sgoedecke/system-one", + "bench_mark": null + }, + { + "id": "zszz3/Pi-Jev-Guide", + "kind": "novel", + "description": "", + "stars": 19, + "why": null, + "html_url": "https://github.com/zszz3/Pi-Jev-Guide", + "bench_mark": null + }, + { + "id": "mithalouni/system-one-open", + "kind": "novel", + "description": "Open replica of TypeSafe's Jev: typed calibrated decisions in one forward pass, on Gemma 4 E2B / Gemma 3 270M (Modal)", + "stars": 18, + "why": null, + "html_url": "https://github.com/mithalouni/system-one-open", + "bench_mark": null + }, + { + "id": "davila7/jev-explained", + "kind": "novel", + "description": "Jev Explained", + "stars": 14, + "why": null, + "html_url": "https://github.com/davila7/jev-explained", + "bench_mark": null + }, + { + "id": "ckaraca/awesome-jev", + "kind": "novel", + "description": "A curated list of tools, integrations, and experiments built on Jev, TypeSafe AI's System One model for fast, typed decisions.", + "stars": 7, + "why": null, + "html_url": "https://github.com/ckaraca/awesome-jev", + "bench_mark": null + }, + { + "id": "arunav25/jev-mcp", + "kind": "novel", + "description": "Connect JEV to MCP clients and compare its judgments against general-purpose LLMs using shared datasets and measurable accuracy.", + "stars": 5, + "why": null, + "html_url": "https://github.com/arunav25/jev-mcp", + "bench_mark": null + }, + { + "id": "cobusgreyling/Jev", + "kind": "novel", + "description": "Unofficial TypeSafe Jev showcase \u2014 System One decisions, not chat.", + "stars": 3, + "why": null, + "html_url": "https://github.com/cobusgreyling/Jev", + "bench_mark": null + }, + { + "id": "nrdz-labs/fast-jev-opencode", + "kind": "novel", + "description": "Jev-scored context pruning for OpenCode: drops stale tool calls and truncates bulky results on the outgoing request \u2014 fail-open, cache-backed, configurable live. Port of fast-jev-compaction to the V2 context hook.", + "stars": 3, + "why": null, + "html_url": "https://github.com/nrdz-labs/fast-jev-opencode", + "bench_mark": null + }, + { + "id": "liao96312/jev-arena-nanojev", + "kind": "novel", + "description": "\u5b8c\u5168\u672c\u5730\u7684 NanoJev \u7f51\u683c\u51b3\u7b56\u6e38\u620f\u5b9e\u9a8c\u573a\uff0c\u652f\u6301\u4e2d\u6587 Pygame\u3001\u591a\u5173\u5361\u4e0e GTX 1660S \u8bad\u7ec3", + "stars": 2, + "why": null, + "html_url": "https://github.com/liao96312/jev-arena-nanojev", + "bench_mark": null + }, + { + "id": "Krug2/JevLM-Open", + "kind": "novel", + "description": "", + "stars": 1, + "why": null, + "html_url": "https://github.com/Krug2/JevLM-Open", + "bench_mark": null + }, + { + "id": "Rizzo-AI-Academy/rizzo-flow", + "kind": "novel", + "description": "The open, local take on Jev: typed decisions from an LLM, without generating a single token", + "stars": 1, + "why": null, + "html_url": "https://github.com/Rizzo-AI-Academy/rizzo-flow", + "bench_mark": null + }, + { + "id": "clouatre-labs/decisions-judge-mcp", + "kind": "novel", + "description": "MCP server exposing an LLM judge (typed decisions: yes/no probability, choice, score) for coding agents", + "stars": 1, + "why": null, + "html_url": "https://github.com/clouatre-labs/decisions-judge-mcp", + "bench_mark": null + }, + { + "id": "emirbartu/opencode-system-one", + "kind": "novel", + "description": "Opencode plugin using Jev (system one model) as part of software development process. Not affiliated with Opencode team.", + "stars": 1, + "why": null, + "html_url": "https://github.com/emirbartu/opencode-system-one", + "bench_mark": null + }, + { + "id": "kotoba-lang/typed-decisions", + "kind": "novel", + "description": "Jev-shaped typed-decision model (state + Choice/Score/Noul questions -> calibrated probabilities, one pass) on ModernBERT / DeBERTa / LLaDA-MoE, with measured latency, accuracy, calibration and training cost", + "stars": 1, + "why": null, + "html_url": "https://github.com/kotoba-lang/typed-decisions", + "bench_mark": null + }, + { + "id": "mallahyari/system-one-benchmark", + "kind": "novel", + "description": "", + "stars": 1, + "why": null, + "html_url": "https://github.com/mallahyari/system-one-benchmark", + "bench_mark": null + }, + { + "id": "sable-inc/jev-linter-action", + "kind": "novel", + "description": "Configurable semantic CI checks for repository files using TypeSafe Jev", + "stars": 1, + "why": null, + "html_url": "https://github.com/sable-inc/jev-linter-action", + "bench_mark": null + }, + { + "id": "shirenchuang/awsomejev", + "kind": "novel", + "description": "Awesome Jev\uff1aJev \u5f00\u6e90\u751f\u6001\u5bfc\u822a", + "stars": 1, + "why": null, + "html_url": "https://github.com/shirenchuang/awsomejev", + "bench_mark": null + }, + { + "id": "1816586742-stack/jev-craft", + "kind": "novel", + "description": "\u8ba9 Agent \u957f\u51fa\u300c\u64cd\u4f5c\u6746\u300d\uff1a\u7528 System One \u6a21\u578b\uff08Jev\uff09\u627f\u62c5\u9ad8\u9891\u5224\u65ad\u3001\u591a\u6a21\u6001\u6a21\u578b\u5f53\u773c\u775b\u3002\u7ed9\u601d\u8def + \u53ef\u8dd1\u7684\u53c2\u8003\u5b9e\u73b0\uff0875 \u6761\u79bb\u7ebf\u65ad\u8a00\uff0c\u96f6\u4f9d\u8d56\u96f6 key\uff09", + "stars": 0, + "why": null, + "html_url": "https://github.com/1816586742-stack/jev-craft", + "bench_mark": null + }, + { + "id": "DolphinMiner/jev-rss", + "kind": "novel", + "description": "A local-first RSS reader with Jev-powered semantic screening. Follow what matters, inspect every judgment, and keep control of your reading. English / \u7b80\u4f53\u4e2d\u6587.", + "stars": 0, + "why": null, + "html_url": "https://github.com/DolphinMiner/jev-rss", + "bench_mark": null + }, + { + "id": "Dreydrey9000/jev-relay", + "kind": "novel", + "description": "Guarded local-first decision advice for Claude Code, Codex and Hermes/Jax, with explicit Jev checks and review fallbacks.", + "stars": 0, + "why": null, + "html_url": "https://github.com/Dreydrey9000/jev-relay", + "bench_mark": null + }, + { + "id": "Eric-Zhou-0302/jev-A-share-trader", + "kind": "novel", + "description": "A Jev-powered technical analysis workspace for China A-shares, supporting AKShare/Tushare, market scans, and Buy/Hold/Sell assessments with time horizons and traceable evidence.", + "stars": 0, + "why": null, + "html_url": "https://github.com/Eric-Zhou-0302/jev-A-share-trader", + "bench_mark": null + }, + { + "id": "JTech-CO/Jev-Simulink-Supervisor", + "kind": "novel", + "description": "Jev-Simulink Supervisor", + "stars": 0, + "why": null, + "html_url": "https://github.com/JTech-CO/Jev-Simulink-Supervisor", + "bench_mark": null + }, + { + "id": "Kwwwww74/OpenJev", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/Kwwwww74/OpenJev", + "bench_mark": null + }, + { + "id": "RuipuCui/jev-harness", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/RuipuCui/jev-harness", + "bench_mark": null + }, + { + "id": "Xubqpanda/JevLoop", + "kind": "novel", + "description": "The agent loop where decisions don't cost a model call. Zero deps, runs offline, no API key needed.", + "stars": 0, + "why": null, + "html_url": "https://github.com/Xubqpanda/JevLoop", + "bench_mark": null + }, + { + "id": "YuanKJing/Jev-as-Policy", + "kind": "novel", + "description": "The highly anticipated open-source repository for JEV as Policy enables one-click setup of the simulation environment. Evaluations of Astra + JEV on benchmarks such as RoboTwin will also be released soon.", + "stars": 0, + "why": null, + "html_url": "https://github.com/YuanKJing/Jev-as-Policy", + "bench_mark": null + }, + { + "id": "Zafer-Liu/jev-demos", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/Zafer-Liu/jev-demos", + "bench_mark": null + }, + { + "id": "aboisvert/jevvy", + "kind": "novel", + "description": "Use jev model to augment csv files with inferred classification, scoring, or probability scores", + "stars": 0, + "why": null, + "html_url": "https://github.com/aboisvert/jevvy", + "bench_mark": null + }, + { + "id": "andrest04/jev-lab", + "kind": "novel", + "description": "Local lab for learning and testing TypeSafe Jev (System One)", + "stars": 0, + "why": null, + "html_url": "https://github.com/andrest04/jev-lab", + "bench_mark": null + }, + { + "id": "andyrewlee/awesome-system-one", + "kind": "novel", + "description": "Curated list of tools related to system one models", + "stars": 0, + "why": null, + "html_url": "https://github.com/andyrewlee/awesome-system-one", + "bench_mark": null + }, + { + "id": "ashafizullah/jev-triage", + "kind": "novel", + "description": "Automated issue & PR triage for open-source maintainers, powered by Jev (TypeSafe AI).", + "stars": 0, + "why": null, + "html_url": "https://github.com/ashafizullah/jev-triage", + "bench_mark": null + }, + { + "id": "blanket11/jev-guide-ja", + "kind": "novel", + "description": "jev\u306e\u5b66\u7fd2", + "stars": 0, + "why": null, + "html_url": "https://github.com/blanket11/jev-guide-ja", + "bench_mark": null + }, + { + "id": "caohy1988/jev-guard-smoke", + "kind": "novel", + "description": "Lab smoke proof for leepokai/jev-guard (Awesome-Jev #1). No API keys.", + "stars": 0, + "why": null, + "html_url": "https://github.com/caohy1988/jev-guard-smoke", + "bench_mark": null + }, + { + "id": "dinkarjuyal/jev-gepa", + "kind": "novel", + "description": "Wiring a fast local NLI judge (Jev) into GEPA's reflective prompt optimization loop", + "stars": 0, + "why": null, + "html_url": "https://github.com/dinkarjuyal/jev-gepa", + "bench_mark": null + }, + { + "id": "early-effect/hexis", + "kind": "novel", + "description": "ZIO / Scala 3 SDK for TypeSafe System One (Jev)", + "stars": 0, + "why": null, + "html_url": "https://github.com/early-effect/hexis", + "bench_mark": null + }, + { + "id": "elberacasa/omawish", + "kind": "novel", + "description": "A System One for Omarchy: type what you want, and your desktop does it. Local, instant, 33M parameters, fine-tuned on one gaming GPU.", + "stars": 0, + "why": null, + "html_url": "https://github.com/elberacasa/omawish", + "bench_mark": null + }, + { + "id": "endman100/research-Qwen3.8-JevLike", + "kind": "novel", + "description": "71-label binary routing vs JSON Schema on Qwen3.8 NVFP4 / RTX 5090: measurements, raw evidence and ideal-parallel analysis", + "stars": 0, + "why": null, + "html_url": "https://github.com/endman100/research-Qwen3.8-JevLike", + "bench_mark": null + }, + { + "id": "eteen12/jev-browser-automation", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/eteen12/jev-browser-automation", + "bench_mark": null + }, + { + "id": "fruitymcdoo/JevChat", + "kind": "novel", + "description": "A chat interface built on Jev, TypeSafe's decision-only model: every word is a typed decision", + "stars": 0, + "why": null, + "html_url": "https://github.com/fruitymcdoo/JevChat", + "bench_mark": null + }, + { + "id": "ismaelsoilet/jev-harness", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/ismaelsoilet/jev-harness", + "bench_mark": null + }, + { + "id": "jacks3tr/Jev-Desktop", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/jacks3tr/Jev-Desktop", + "bench_mark": null + }, + { + "id": "javsanesq/jevlab", + "kind": "novel", + "description": "A terminal workbench for learning, testing, and connecting TypeSafe Jev decisions", + "stars": 0, + "why": null, + "html_url": "https://github.com/javsanesq/jevlab", + "bench_mark": null + }, + { + "id": "joelakaufmann-lgtm/NRS-Navigator", + "kind": "novel", + "description": "Local Nevada statute search and a reproducible evaluation of Jev-assisted ranking against keyword search.", + "stars": 0, + "why": null, + "html_url": "https://github.com/joelakaufmann-lgtm/NRS-Navigator", + "bench_mark": null + }, + { + "id": "k-srkw/jev-playground", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/k-srkw/jev-playground", + "bench_mark": null + }, + { + "id": "koteitan/laya-bot-det", + "kind": "novel", + "description": "bot detector by laya for nostr", + "stars": 0, + "why": null, + "html_url": "https://github.com/koteitan/laya-bot-det", + "bench_mark": null + }, + { + "id": "kzkhykw/jev-or-not", + "kind": "novel", + "description": "Jev\u308b\uff1f\u30e2\u30c7\u30eb\u9078\u629e\u30d5\u30ed\u30fc\u30c1\u30e3\u30fc\u30c8", + "stars": 0, + "why": null, + "html_url": "https://github.com/kzkhykw/jev-or-not", + "bench_mark": null + }, + { + "id": "laidick/system-one-benchmark", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/laidick/system-one-benchmark", + "bench_mark": null + }, + { + "id": "lezgoverci/jev-docs", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/lezgoverci/jev-docs", + "bench_mark": null + }, + { + "id": "luckberonne/mini-jev", + "kind": "novel", + "description": "Clasificador de comandos de shell de una sola pasada (solo lectura / reversible / destructivo), inspirado en Jev", + "stars": 0, + "why": null, + "html_url": "https://github.com/luckberonne/mini-jev", + "bench_mark": null + }, + { + "id": "marcosmartinez/jev-acento", + "kind": "novel", + "description": "\u00bfJev entiende tu acento? Pre-registered audit of TypeSafe AI's Jev on Spanish \u2014 accuracy, calibration and token cost \u2014 plus a CLI to run the same comparison on your own labelled data.", + "stars": 0, + "why": null, + "html_url": "https://github.com/marcosmartinez/jev-acento", + "bench_mark": null + }, + { + "id": "mednabouli/jev-ai-polymarket-copy-trading", + "kind": "novel", + "description": "Automated Polymarket copy trading bot with MCP servers, Telegram alerts, and profitable wallet tracking. Zero API keys - uses Claude Code OAuth session auth.", + "stars": 0, + "why": null, + "html_url": "https://github.com/mednabouli/jev-ai-polymarket-copy-trading", + "bench_mark": null + }, + { + "id": "olivdx/jev-mcp", + "kind": "novel", + "description": "Jev-powered decision layer for coding agents. Analyze code and diffs, assess bugs, security, risk, and breaking changes, and return structured decisions for automated continue, fix, retry, or human-review workflows.", + "stars": 0, + "why": null, + "html_url": "https://github.com/olivdx/jev-mcp", + "bench_mark": null + }, + { + "id": "pattoor/JEV-agent-opencv", + "kind": "novel", + "description": "Juego simple para probar el modelo JEV con vision", + "stars": 0, + "why": null, + "html_url": "https://github.com/pattoor/JEV-agent-opencv", + "bench_mark": null + }, + { + "id": "peach-zhang/typesafe-go", + "kind": "novel", + "description": "TypeSafe System One (Jev) \u7684 Go SDK \u2014 \u7c7b\u578b\u5316\u5224\u65ad\u4e0e\u6982\u7387,\u4ee3\u7801\u638c\u63a7\u5de5\u4f5c\u6d41", + "stars": 0, + "why": null, + "html_url": "https://github.com/peach-zhang/typesafe-go", + "bench_mark": null + }, + { + "id": "promptgtm-shared/clay-jev-people-ranker", + "kind": "novel", + "description": "Agent Skill and Python workflow for Clay lead scoring, B2B prospect qualification, and people-search ranking with TypeSafe JEV.", + "stars": 0, + "why": null, + "html_url": "https://github.com/promptgtm-shared/clay-jev-people-ranker", + "bench_mark": null + }, + { + "id": "sahasrarjn/system-one", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/sahasrarjn/system-one", + "bench_mark": null + }, + { + "id": "sarathi-aiml/jevsql", + "kind": "novel", + "description": "Text-to-SQL where the model never writes SQL \u2014 typed, calibrated decisions (TypeSafe Jev) + code-assembled queries", + "stars": 0, + "why": null, + "html_url": "https://github.com/sarathi-aiml/jevsql", + "bench_mark": null + }, + { + "id": "theosunny/jev_stock", + "kind": "novel", + "description": "", + "stars": 0, + "why": null, + "html_url": "https://github.com/theosunny/jev_stock", + "bench_mark": null + }, + { + "id": "twilwa/pi-typesafe", + "kind": "novel", + "description": "Pi coding-agent extension built on the TypeSafe AI System One API (Jev)", + "stars": 0, + "why": null, + "html_url": "https://github.com/twilwa/pi-typesafe", + "bench_mark": null + }, + { + "id": "willgriffin/pi-fusion-matrix", + "kind": "novel", + "description": "Multi-model deliberation for the pi coding agent: named fusions, per-slot fallback and routing, version-free aliases, conservative decision backends", + "stars": 0, + "why": null, + "html_url": "https://github.com/willgriffin/pi-fusion-matrix", + "bench_mark": null + }, + { + "id": "Zefan-Cai/Open-Jev", + "kind": "revisit", + "description": "REVISIT densify: HEAD 4933ee84951f (prior pushed 00:43\u219201:34); archived published benchmark thread + latency chart; recent commits publish audited Astra/Luna/TREC results (*theirs*)", + "stars": 4, + "why": "REVISIT densify: HEAD 4933ee84951f (prior pushed 00:43\u219201:34); archived published benchmark thread + latency chart; recent commits publish audited Astra/Luna/TREC results (*theirs*)", + "html_url": "https://github.com/Zefan-Cai/Open-Jev", + "bench_mark": "*theirs*" + }, + { + "id": "alexwestco/llm-to-jev", + "kind": "revisit", + "description": "Convert LLM prompts to Jev prompts", + "stars": 3, + "why": "REVISIT: description rewrite 'llm-to-jev' \u2192 'Convert LLM prompts to Jev prompts' (desc_hash change); SHA unchanged 234058ab372d", + "html_url": "https://github.com/alexwestco/llm-to-jev", + "bench_mark": null + } +] \ No newline at end of file diff --git a/research/archive/hourly/2026-09-21T01/github.json b/research/archive/hourly/2026-09-21T01/github.json new file mode 100644 index 0000000..ae177cc --- /dev/null +++ b/research/archive/hourly/2026-09-21T01/github.json @@ -0,0 +1,1768 @@ +[ + { + "id": "brainstormity/Jev-X-Sentiment-Analysis", + "description": null, + "stars": 136, + "forks": 20, + "license": null, + "language": "Python", + "size": 467, + "pushed_at": "2026-09-19T22:51:11Z", + "created_at": "2026-09-19T22:51:07Z", + "updated_at": "2026-09-21T01:57:49Z", + "homepage": null, + "default_branch": "main", + "default_sha": "5c932f941a92348781e8e5b471a9ddb6af980253", + "readme": { + "path": "README.md", + "sha": "bf4134b44cda73599f005e9408391353a2ed9437", + "size": 7370 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "ikermoel/open-alternative-jev", + "description": "Open-source alternative to TypeSafe's Jev: a System One style model layer that gives typed, calibrated decisions from any open-weights LLM in one forward pass (HF + vLLM), with honest benchmarks", + "stars": 36, + "forks": 8, + "license": "Apache-2.0", + "language": "Python", + "size": 2025, + "pushed_at": "2026-09-20T03:14:39Z", + "created_at": "2026-09-18T05:20:56Z", + "updated_at": "2026-09-20T23:48:09Z", + "homepage": "https://huggingface.co/spaces/IkerMoel/open-alternative-jev", + "default_branch": "main", + "default_sha": "6ad87d7ce2f4ef472acb9253419134a2643a57d2", + "readme": { + "path": "README.md", + "sha": "d4653d9d1b95a0d584ea1fb9c7e11426463d7fd6", + "size": 11764 + }, + "topics": [ + "calibration", + "classification", + "jev", + "jev-alternative", + "llm", + "logprobs", + "open-jev", + "open-source-jev", + "open-weights", + "qwen", + "structured-decisions", + "structured-output", + "system-one", + "system-one-model", + "transformers", + "typed-decisions", + "typesafe", + "typesafe-jev", + "vllm" + ], + "desc_hash": "3aeedb10b4e9" + }, + { + "id": "okinaaudio/live-jev", + "description": "Control Ableton Live with one short sentence (Japanese / English). Summon with \u2318\u21e7Space, type or dictate, done.", + "stars": 36, + "forks": 2, + "license": "MIT", + "language": "Python", + "size": 522, + "pushed_at": "2026-09-21T01:46:35Z", + "created_at": "2026-09-19T03:15:09Z", + "updated_at": "2026-09-21T01:46:40Z", + "homepage": null, + "default_branch": "main", + "default_sha": "2446eb777ad9f59f77b96ee5b081ee8c2812e0a3", + "readme": { + "path": "README.md", + "sha": "52e7d45c799f9b4d20053d2c7e28b8caac0e1b07", + "size": 12968 + }, + "topics": [ + "ableton-live", + "macos", + "music-production", + "natural-language" + ], + "desc_hash": "73f6f5c579e5" + }, + { + "id": "heyjunpenn/awesome-jev", + "description": "A verified, community-maintained catalog of 503 open-source projects built with Jev.", + "stars": 32, + "forks": 2, + "license": "MIT", + "language": "TypeScript", + "size": 3326, + "pushed_at": "2026-09-20T10:08:27Z", + "created_at": "2026-09-19T05:14:36Z", + "updated_at": "2026-09-21T01:47:12Z", + "homepage": "https://jevbest.com/", + "default_branch": "main", + "default_sha": "8ecdef6a6fcb3acc0c498fe8a87855e5bcd21c0f", + "readme": { + "path": "README.md", + "sha": "0b24293bc84ff0cc828b67c7d5049b74d0d085ed", + "size": 104955 + }, + "topics": [], + "desc_hash": "eed807328057" + }, + { + "id": "NanmiCoder/jev-arena", + "description": "Jev \u6a21\u578b\u4ecb\u7ecd\u4e0e\u5b9e\u6d4b\uff1a\u901a\u8fc7 Choice / Score / Noul \u5c06\u81ea\u7136\u8bed\u8a00\u8f6c\u4e3a\u5e26\u7c7b\u578b\u7684\u5224\u65ad\u4e0e\u6982\u7387\uff0c\u7528\u4e8e\u5206\u7c7b\u3001\u8bc4\u5206\u548c\u8def\u7531\uff1b\u652f\u6301\u4e0e DeepSeek \u7b49\u6a21\u578b\u5bf9\u6bd4\u8bc4\u8bba\u6253\u6807\u3001\u901f\u5ea6\u4e0e\u7ed3\u679c\uff0c\u542b CSV/Excel \u5bfc\u5165\u3001\u539f\u901f\u56de\u653e\u4e0e\u79bb\u7ebf\u62a5\u544a\u3002", + "stars": 31, + "forks": 3, + "license": "MIT", + "language": "JavaScript", + "size": 18427, + "pushed_at": "2026-09-20T15:00:39Z", + "created_at": "2026-09-19T11:01:19Z", + "updated_at": "2026-09-21T01:37:34Z", + "homepage": "https://nanmicoder.github.io/jev-arena/", + "default_branch": "main", + "default_sha": "2ca160cc4aa9ac72a4341e2ac5903258e8c69c84", + "readme": { + "path": "README.md", + "sha": "4eb7f2dec20a2ecdf0d74a6254f311ee8bff19a7", + "size": 2878 + }, + "topics": [], + "desc_hash": "e2732659dccb" + }, + { + "id": "yzfly/awesome-jev-zh", + "description": "Jev / TypeSafe System One \u4e2d\u6587\u7cbe\u9009\u5217\u8868\uff1a\u5b98\u65b9\u8d44\u6599\u3001SDK\u3001\u7206\u6b3e\u5e94\u7528\u3001Agent \u5de5\u5177\u3001\u5f00\u6e90\u590d\u73b0\u4e0e\u72ec\u7acb\u8bc4\u6d4b\uff0c\u9644\u4e2d\u6587\u4e0a\u624b\u6307\u5357\uff0c\u6bcf\u65e5\u81ea\u52a8\u6536\u5f55 GitHub \u70ed\u95e8\u9879\u76ee\u3002", + "stars": 31, + "forks": 9, + "license": "CC0-1.0", + "language": "HTML", + "size": 515, + "pushed_at": "2026-09-21T01:20:52Z", + "created_at": "2026-09-18T06:33:02Z", + "updated_at": "2026-09-21T01:41:00Z", + "homepage": "https://code.jiangshu.ai/awesome-jev-zh/", + "default_branch": "main", + "default_sha": "7cce63fedfc507993f311ced0f270e1a2739695a", + "readme": { + "path": "README.md", + "sha": "f8c549471e6ae4bbc3525408fda7c7e1d1feb626", + "size": 106685 + }, + "topics": [ + "ai-agent", + "awesome", + "awesome-list", + "chinese", + "jev", + "llm", + "structured-output", + "system-one", + "typesafe", + "typesafe-ai" + ], + "desc_hash": "1ee1f7288a28" + }, + { + "id": "openroboto-ai/jev-robot-control", + "description": null, + "stars": 27, + "forks": 0, + "license": "NOASSERTION", + "language": "Python", + "size": 18965, + "pushed_at": "2026-09-19T13:39:07Z", + "created_at": "2026-09-19T13:35:41Z", + "updated_at": "2026-09-20T20:15:14Z", + "homepage": null, + "default_branch": "main", + "default_sha": "7a4ed8b72c3c17d7aa790678ed9660df67c10dd3", + "readme": { + "path": "README.md", + "sha": "0e749c38197e78c46dacbb1f61910ef2b244a823", + "size": 8306 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "win4r/jev-skill-suggester", + "description": "\u7528 TypeSafe Jev \u63a8\u8350\u5df2\u5b89\u88c5 Skill / Bounded installed-skill recommendations with TypeSafe Jev. Python CLI, Codex skill, bilingual docs and live examples.", + "stars": 27, + "forks": 2, + "license": "MIT", + "language": "Python", + "size": 47, + "pushed_at": "2026-09-19T15:31:40Z", + "created_at": "2026-09-19T15:31:35Z", + "updated_at": "2026-09-21T00:32:21Z", + "homepage": null, + "default_branch": "main", + "default_sha": "05fbd7ce9ec74a2f193c09276a8d4d077d9d5e5c", + "readme": { + "path": "README.md", + "sha": "2a03bec83f3c9915d7c8691523a7a489c65fee7e", + "size": 11404 + }, + "topics": [], + "desc_hash": "a146a470c049" + }, + { + "id": "PyModel/typesafe-mcp", + "description": null, + "stars": 23, + "forks": 2, + "license": "MIT", + "language": "Go", + "size": 1123, + "pushed_at": "2026-09-21T01:48:45Z", + "created_at": "2026-09-19T18:43:24Z", + "updated_at": "2026-09-21T01:48:36Z", + "homepage": null, + "default_branch": "main", + "default_sha": "cf01808bf27485a017cc6eebead6384556dab552", + "readme": { + "path": "README.md", + "sha": "c7c36760aa1cfb1c4db3bafdc867b4bafd1dd241", + "size": 15857 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "AkashPriyadarshii/jev-seo", + "description": "100% free \u20b90 agent-first SEO & GEO CLI suite and MCP server in Rust replacing Semrush and OpenSEO via DuckDuckGo and TypeSafe Jev System One", + "stars": 21, + "forks": 1, + "license": "MIT", + "language": "Rust", + "size": 102, + "pushed_at": "2026-09-21T01:51:33Z", + "created_at": "2026-09-18T13:02:32Z", + "updated_at": "2026-09-21T01:50:37Z", + "homepage": "https://jevseo.vercel.app", + "default_branch": "master", + "default_sha": "f8cb7c55c356cbc12929cf548ae75ea568f03861", + "readme": { + "path": "README.md", + "sha": "e3290fd15adde74b7081fca7503e65e228306419", + "size": 11529 + }, + "topics": [ + "ahrefs-alternative", + "claude-code", + "cli", + "decision-oracle", + "duckduckgo", + "generative-engine-optimization", + "geo", + "jev", + "mcp", + "mcp-server", + "rank-tracker", + "rust", + "search-engine-optimization", + "semrush-alternative", + "seo", + "seo-audit", + "seo-tools", + "serp", + "typesafe", + "typesafe-ai" + ], + "desc_hash": "3ec093c905b3" + }, + { + "id": "devtooligan/jevscan-evm", + "description": null, + "stars": 21, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 575, + "pushed_at": "2026-09-19T20:15:42Z", + "created_at": "2026-09-19T00:25:35Z", + "updated_at": "2026-09-20T22:02:31Z", + "homepage": null, + "default_branch": "main", + "default_sha": "2ecbe7cd93a92adb4464821d3ca7b0828c0c1b3c", + "readme": { + "path": "README.md", + "sha": "075a01007b92aec41fd2a4ff52d1c69d07840028", + "size": 8763 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "mizzlelover/jev-hub", + "description": "JEV HUB \u00b7 X \u4e0a\u5173\u4e8e TypeSafe AI\u300c\u7cfb\u7edf\u4e00\u6a21\u578b\u300dJev \u7684\u957f\u6587\u4e0e\u6f14\u793a\u89c6\u9891\u805a\u5408\uff08\u4fdd\u7559\u539f\u94fe\u4e0e\u4f5c\u8005\uff09\uff5c \u8c01\u662f\u4e13\u5bb6 \u51fa\u54c1", + "stars": 21, + "forks": 3, + "license": "NOASSERTION", + "language": "CSS", + "size": 189, + "pushed_at": "2026-09-19T02:54:57Z", + "created_at": "2026-09-19T02:32:02Z", + "updated_at": "2026-09-20T07:21:23Z", + "homepage": "https://mizzlelover.github.io/jev-hub/", + "default_branch": "main", + "default_sha": "4262d71d2d06a910349d709b4d4958ce34c5ed98", + "readme": { + "path": "README.md", + "sha": "04530e25b75136c90c0e65569dc13fb07a55aab1", + "size": 55472 + }, + "topics": [ + "ai", + "awesome-list", + "jev", + "typesafe", + "x-twitter" + ], + "desc_hash": "a5b7621eaac0" + }, + { + "id": "sgoedecke/system-one", + "description": "Batched single-token choice inference for open language models, compatible with TypeSafe", + "stars": 20, + "forks": 3, + "license": null, + "language": "Python", + "size": 125589, + "pushed_at": "2026-09-18T03:22:07Z", + "created_at": "2026-09-17T12:13:54Z", + "updated_at": "2026-09-20T21:23:45Z", + "homepage": null, + "default_branch": "main", + "default_sha": "ebde2a2db7067b920dfe51e9ce785613e66613d5", + "readme": { + "path": "README.md", + "sha": "d331b567e2c37146a6194004ee494d5f9bbc345d", + "size": 3331 + }, + "topics": [], + "desc_hash": "88c6ac90a44e" + }, + { + "id": "zszz3/Pi-Jev-Guide", + "description": null, + "stars": 19, + "forks": 0, + "license": "MIT", + "language": "TypeScript", + "size": 82, + "pushed_at": "2026-09-20T03:41:56Z", + "created_at": "2026-09-19T09:22:39Z", + "updated_at": "2026-09-20T08:42:41Z", + "homepage": null, + "default_branch": "main", + "default_sha": "f187d464062bfd6fb41ae690c15ad1af974cbfb6", + "readme": { + "path": "README.md", + "sha": "45aaa093f82bf3cd81b93451e7b06d9bbbfd06d1", + "size": 14302 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "mithalouni/system-one-open", + "description": "Open replica of TypeSafe's Jev: typed calibrated decisions in one forward pass, on Gemma 4 E2B / Gemma 3 270M (Modal)", + "stars": 18, + "forks": 4, + "license": "NOASSERTION", + "language": "Python", + "size": 13272, + "pushed_at": "2026-09-17T07:06:55Z", + "created_at": "2026-09-17T05:56:58Z", + "updated_at": "2026-09-20T18:12:14Z", + "homepage": null, + "default_branch": "main", + "default_sha": "77f1f7cccf8aa752e0ed7edcc8d2094bac707bdc", + "readme": { + "path": "README.md", + "sha": "535f33028a685b909fbc287409697be55da6756a", + "size": 5511 + }, + "topics": [], + "desc_hash": "8755f1023081" + }, + { + "id": "davila7/jev-explained", + "description": "Jev Explained", + "stars": 14, + "forks": 1, + "license": "MIT", + "language": "TypeScript", + "size": 1180, + "pushed_at": "2026-09-20T16:53:26Z", + "created_at": "2026-09-19T11:46:20Z", + "updated_at": "2026-09-21T00:04:33Z", + "homepage": "https://jev-explained-drab.vercel.app", + "default_branch": "main", + "default_sha": "5cbe35e04609112be77b1bd447bd79b3bde7980b", + "readme": { + "path": "README.md", + "sha": "7c69d941dc821e2df3a37e3348357f04b0b2adb8", + "size": 5548 + }, + "topics": [], + "desc_hash": "a20ee63f34c9" + }, + { + "id": "ckaraca/awesome-jev", + "description": "A curated list of tools, integrations, and experiments built on Jev, TypeSafe AI's System One model for fast, typed decisions.", + "stars": 7, + "forks": 1, + "license": "CC0-1.0", + "language": "Python", + "size": 47, + "pushed_at": "2026-09-20T13:48:37Z", + "created_at": "2026-09-19T00:52:55Z", + "updated_at": "2026-09-20T13:48:41Z", + "homepage": "https://docs.typesafe.ai/", + "default_branch": "main", + "default_sha": "2207d19d91fa80d54b4de3f86295437605be58e8", + "readme": { + "path": "README.md", + "sha": "3e58af61ccebd902207cc784eee84b8318d027d1", + "size": 20131 + }, + "topics": [ + "ai-agents", + "awesome", + "awesome-list", + "computer-use", + "jev", + "llm", + "system-one", + "typesafe" + ], + "desc_hash": "70b1ba2dc046" + }, + { + "id": "arunav25/jev-mcp", + "description": "Connect JEV to MCP clients and compare its judgments against general-purpose LLMs using shared datasets and measurable accuracy.", + "stars": 5, + "forks": 0, + "license": "MIT", + "language": "JavaScript", + "size": 69, + "pushed_at": "2026-09-17T16:05:17Z", + "created_at": "2026-09-17T15:19:03Z", + "updated_at": "2026-09-18T16:03:38Z", + "homepage": null, + "default_branch": "main", + "default_sha": "036d32433c0d234d37ef11a6a44619b69444556c", + "readme": { + "path": "README.md", + "sha": "a858bf2227da3012651457b31404f70f08fb807f", + "size": 10859 + }, + "topics": [], + "desc_hash": "adbf6ad85063" + }, + { + "id": "cobusgreyling/Jev", + "description": "Unofficial TypeSafe Jev showcase \u2014 System One decisions, not chat.", + "stars": 3, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 347, + "pushed_at": "2026-09-20T07:32:08Z", + "created_at": "2026-09-19T10:35:48Z", + "updated_at": "2026-09-20T07:32:11Z", + "homepage": "https://docs.typesafe.ai/", + "default_branch": "main", + "default_sha": "c636087edf4100a097a456e559c4b2fa769b80e8", + "readme": { + "path": "README.md", + "sha": "9de601117711b440fc64776052511a2dbc099240", + "size": 9262 + }, + "topics": [ + "calibrated-decisions", + "jev", + "showcase", + "system-one", + "typesafe" + ], + "desc_hash": "086eb9b31d04" + }, + { + "id": "nrdz-labs/fast-jev-opencode", + "description": "Jev-scored context pruning for OpenCode: drops stale tool calls and truncates bulky results on the outgoing request \u2014 fail-open, cache-backed, configurable live. Port of fast-jev-compaction to the V2 context hook.", + "stars": 3, + "forks": 0, + "license": "MIT", + "language": "TypeScript", + "size": 140, + "pushed_at": "2026-09-19T06:25:17Z", + "created_at": "2026-09-19T03:51:47Z", + "updated_at": "2026-09-20T22:59:18Z", + "homepage": null, + "default_branch": "main", + "default_sha": "4c4de4dcffa0201502fcb682c80b807cfd248140", + "readme": { + "path": "README.md", + "sha": "1a733604c7fb0757ae2e72e15c1829ef22deb366", + "size": 3814 + }, + "topics": [ + "compaction", + "context", + "jev", + "opencode", + "opencode-plugin", + "typesafe" + ], + "desc_hash": "50253fd57e4c" + }, + { + "id": "liao96312/jev-arena-nanojev", + "description": "\u5b8c\u5168\u672c\u5730\u7684 NanoJev \u7f51\u683c\u51b3\u7b56\u6e38\u620f\u5b9e\u9a8c\u573a\uff0c\u652f\u6301\u4e2d\u6587 Pygame\u3001\u591a\u5173\u5361\u4e0e GTX 1660S \u8bad\u7ec3", + "stars": 2, + "forks": 0, + "license": null, + "language": "Python", + "size": 13967, + "pushed_at": "2026-09-21T01:46:57Z", + "created_at": "2026-09-19T02:23:40Z", + "updated_at": "2026-09-21T01:47:01Z", + "homepage": null, + "default_branch": "main", + "default_sha": "a4d2ec811fc48d067a3c68ee9e976a2ee7d9a488", + "readme": { + "path": "README.md", + "sha": "a5887c4c76ff032ba3e5b7339bc12fc31e719f05", + "size": 13199 + }, + "topics": [], + "desc_hash": "0ceaea47dd48" + }, + { + "id": "Krug2/JevLM-Open", + "description": null, + "stars": 1, + "forks": 2, + "license": null, + "language": "Python", + "size": 10170, + "pushed_at": "2026-09-20T21:17:19Z", + "created_at": "2026-09-20T18:35:30Z", + "updated_at": "2026-09-21T00:44:46Z", + "homepage": null, + "default_branch": "main", + "default_sha": "84272bdabbea7ed76548b4fdd7d623a9373d01de", + "readme": null, + "topics": [], + "desc_hash": null + }, + { + "id": "Rizzo-AI-Academy/rizzo-flow", + "description": "The open, local take on Jev: typed decisions from an LLM, without generating a single token", + "stars": 1, + "forks": 0, + "license": "Apache-2.0", + "language": "Python", + "size": 1767, + "pushed_at": "2026-09-21T01:55:02Z", + "created_at": "2026-09-21T00:00:02Z", + "updated_at": "2026-09-21T01:55:06Z", + "homepage": "", + "default_branch": "main", + "default_sha": "d97ef676a40e2ad47541835f8676b4be60331911", + "readme": { + "path": "README.md", + "sha": "529432bebd5f6692f1ccffe6bbb0176dbc446aa9", + "size": 23362 + }, + "topics": [], + "desc_hash": "d006ffab3676" + }, + { + "id": "clouatre-labs/decisions-judge-mcp", + "description": "MCP server exposing an LLM judge (typed decisions: yes/no probability, choice, score) for coding agents", + "stars": 1, + "forks": 0, + "license": "Apache-2.0", + "language": "JavaScript", + "size": 114, + "pushed_at": "2026-09-21T01:49:37Z", + "created_at": "2026-09-20T22:01:23Z", + "updated_at": "2026-09-21T01:49:39Z", + "homepage": null, + "default_branch": "main", + "default_sha": "b6fff8ee714cfbbe94c6d9fc09dadfefd57a6676", + "readme": { + "path": "README.md", + "sha": "cd771f8872c011937bd41efdbf275669cfc0e6e8", + "size": 3383 + }, + "topics": [ + "ai", + "cli", + "decisions", + "judge", + "llm", + "mcp", + "mcp-server", + "model-context-protocol" + ], + "desc_hash": "79d7cb28cf57" + }, + { + "id": "emirbartu/opencode-system-one", + "description": "Opencode plugin using Jev (system one model) as part of software development process. Not affiliated with Opencode team.", + "stars": 1, + "forks": 0, + "license": null, + "language": "TypeScript", + "size": 70, + "pushed_at": "2026-09-19T00:28:05Z", + "created_at": "2026-09-19T00:28:02Z", + "updated_at": "2026-09-19T01:30:22Z", + "homepage": "", + "default_branch": "main", + "default_sha": "e51f5fde5ed235004970c3a90fd98033c3dcfdb5", + "readme": { + "path": "README.md", + "sha": "3c26ede85bd1644eca0787085d68af0286f085f2", + "size": 2910 + }, + "topics": [ + "jev", + "opencode-plugin", + "opencode-plugins" + ], + "desc_hash": "791b87501e61" + }, + { + "id": "kotoba-lang/typed-decisions", + "description": "Jev-shaped typed-decision model (state + Choice/Score/Noul questions -> calibrated probabilities, one pass) on ModernBERT / DeBERTa / LLaDA-MoE, with measured latency, accuracy, calibration and training cost", + "stars": 1, + "forks": 0, + "license": "NOASSERTION", + "language": "Python", + "size": 14217, + "pushed_at": "2026-09-20T02:58:24Z", + "created_at": "2026-09-18T07:03:52Z", + "updated_at": "2026-09-20T02:57:55Z", + "homepage": null, + "default_branch": "main", + "default_sha": "10d7834d3b99041f890db4615fb38ef95ced50cc", + "readme": { + "path": "README.md", + "sha": "4d6bbf4c4e446270dea99bb8b9d70bf2df628f09", + "size": 55530 + }, + "topics": [], + "desc_hash": "8588e6be9350" + }, + { + "id": "mallahyari/system-one-benchmark", + "description": null, + "stars": 1, + "forks": 0, + "license": null, + "language": "Python", + "size": 48, + "pushed_at": "2026-09-19T15:55:09Z", + "created_at": "2026-09-19T03:35:00Z", + "updated_at": "2026-09-19T15:55:12Z", + "homepage": null, + "default_branch": "main", + "default_sha": "46c232b887d41b4b560a1eac74972eade3b1fddc", + "readme": { + "path": "README.md", + "sha": "b4b8e76713c89ee429923a13c25577dd5c53f8cf", + "size": 6836 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "sable-inc/jev-linter-action", + "description": "Configurable semantic CI checks for repository files using TypeSafe Jev", + "stars": 1, + "forks": 0, + "license": "MIT", + "language": "JavaScript", + "size": 64, + "pushed_at": "2026-09-21T01:43:28Z", + "created_at": "2026-09-21T00:02:01Z", + "updated_at": "2026-09-21T01:48:23Z", + "homepage": null, + "default_branch": "main", + "default_sha": "1f6ba701fe72a70c2ec070c8eac3dff836a0d704", + "readme": { + "path": "README.md", + "sha": "b1ac97554b8269cacca8c0989cf402aa235e593d", + "size": 4791 + }, + "topics": [], + "desc_hash": "84b832b28033" + }, + { + "id": "shirenchuang/awsomejev", + "description": "Awesome Jev\uff1aJev \u5f00\u6e90\u751f\u6001\u5bfc\u822a", + "stars": 1, + "forks": 0, + "license": "MIT", + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:34:16Z", + "created_at": "2026-09-21T01:25:12Z", + "updated_at": "2026-09-21T01:44:17Z", + "homepage": null, + "default_branch": "main", + "default_sha": "a994099f0a32d027b7229fedf560eaf7a2bb3c46", + "readme": { + "path": "README.md", + "sha": "96a2075be5bb611ac82986ab611cea0806749c16", + "size": 52785 + }, + "topics": [], + "desc_hash": "c4be2e0491a3" + }, + { + "id": "1816586742-stack/jev-craft", + "description": "\u8ba9 Agent \u957f\u51fa\u300c\u64cd\u4f5c\u6746\u300d\uff1a\u7528 System One \u6a21\u578b\uff08Jev\uff09\u627f\u62c5\u9ad8\u9891\u5224\u65ad\u3001\u591a\u6a21\u6001\u6a21\u578b\u5f53\u773c\u775b\u3002\u7ed9\u601d\u8def + \u53ef\u8dd1\u7684\u53c2\u8003\u5b9e\u73b0\uff0875 \u6761\u79bb\u7ebf\u65ad\u8a00\uff0c\u96f6\u4f9d\u8d56\u96f6 key\uff09", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "JavaScript", + "size": 0, + "pushed_at": "2026-09-21T01:20:50Z", + "created_at": "2026-09-21T01:11:38Z", + "updated_at": "2026-09-21T01:20:54Z", + "homepage": null, + "default_branch": "main", + "default_sha": "ad4926f663b8c382f6db7e889288747ccf68d119", + "readme": { + "path": "README.md", + "sha": "b614d4f0b3133685b56d87d2eab4570abdedb92e", + "size": 13263 + }, + "topics": [ + "agent", + "ai-agent", + "decision-making", + "deepseek", + "dsh", + "jev", + "llm", + "minecraft", + "system-one", + "typesafe" + ], + "desc_hash": "5184161785c9" + }, + { + "id": "DolphinMiner/jev-rss", + "description": "A local-first RSS reader with Jev-powered semantic screening. Follow what matters, inspect every judgment, and keep control of your reading. English / \u7b80\u4f53\u4e2d\u6587.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "TypeScript", + "size": 0, + "pushed_at": "2026-09-21T01:03:30Z", + "created_at": "2026-09-21T01:02:51Z", + "updated_at": "2026-09-21T01:03:57Z", + "homepage": "", + "default_branch": "main", + "default_sha": "4a2dbb770faa3a9cac80d7b1f27d181f25b5bb91", + "readme": { + "path": "README.md", + "sha": "caaf2f1ebb6a62052ee95d268c16fb568aa58838", + "size": 8265 + }, + "topics": [ + "ai", + "jev", + "local-first", + "react", + "rss", + "rss-reader", + "rsshub", + "self-hosted", + "semantic-filtering", + "typescript" + ], + "desc_hash": "61001e76fee3" + }, + { + "id": "Dreydrey9000/jev-relay", + "description": "Guarded local-first decision advice for Claude Code, Codex and Hermes/Jax, with explicit Jev checks and review fallbacks.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 5213, + "pushed_at": "2026-09-21T00:59:23Z", + "created_at": "2026-09-21T00:47:38Z", + "updated_at": "2026-09-21T00:59:27Z", + "homepage": null, + "default_branch": "main", + "default_sha": "d52ffe1fc023ff607458c1aee1eba05d1f628f1e", + "readme": { + "path": "README.md", + "sha": "2039013eb53d6098a6c2dd9ecb2c46b52c37462a", + "size": 6375 + }, + "topics": [], + "desc_hash": "dff46c81ca68" + }, + { + "id": "Eric-Zhou-0302/jev-A-share-trader", + "description": "A Jev-powered technical analysis workspace for China A-shares, supporting AKShare/Tushare, market scans, and Buy/Hold/Sell assessments with time horizons and traceable evidence.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 217, + "pushed_at": "2026-09-21T01:43:35Z", + "created_at": "2026-09-20T04:53:28Z", + "updated_at": "2026-09-21T01:43:39Z", + "homepage": null, + "default_branch": "main", + "default_sha": "a7ad82306fbe8e0a2ef92a54f236684c7e4139d5", + "readme": { + "path": "README.md", + "sha": "9c1cd8d69c85c0e71c7159dfdcaefaab5ce4ed7b", + "size": 9451 + }, + "topics": [ + "a-shares", + "akshare", + "fastapi", + "jev", + "python", + "react", + "stock-analysis", + "system-one", + "technical-analysis", + "tushare", + "typesafe-ai" + ], + "desc_hash": "09fad6bc411d" + }, + { + "id": "JTech-CO/Jev-Simulink-Supervisor", + "description": "Jev-Simulink Supervisor", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "MATLAB", + "size": 0, + "pushed_at": "2026-09-21T01:54:08Z", + "created_at": "2026-09-21T01:15:44Z", + "updated_at": "2026-09-21T01:54:11Z", + "homepage": null, + "default_branch": "main", + "default_sha": "76bc14a89694a2eb275239a96e8df6044c18d6d2", + "readme": { + "path": "README.md", + "sha": "61f40464fa61b7b108e0a7dd5b48257e5dca70c3", + "size": 3211 + }, + "topics": [], + "desc_hash": "9e348cc4366c" + }, + { + "id": "Kwwwww74/OpenJev", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:31:34Z", + "created_at": "2026-09-21T01:31:33Z", + "updated_at": "2026-09-21T01:31:37Z", + "homepage": null, + "default_branch": "main", + "default_sha": "423875b6088bd2d9b4e639777e2bdafdd60c760e", + "readme": { + "path": "README.md", + "sha": "8f5ec58f92ea2ad2dddf9ad9e03e3153e38dc4a7", + "size": 9 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "RuipuCui/jev-harness", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T00:02:26Z", + "created_at": "2026-09-21T00:02:25Z", + "updated_at": "2026-09-21T00:02:26Z", + "homepage": null, + "default_branch": "main", + "default_sha": null, + "readme": null, + "topics": [], + "desc_hash": null + }, + { + "id": "Xubqpanda/JevLoop", + "description": "The agent loop where decisions don't cost a model call. Zero deps, runs offline, no API key needed.", + "stars": 0, + "forks": 0, + "license": "Apache-2.0", + "language": "TypeScript", + "size": 150, + "pushed_at": "2026-09-21T01:46:46Z", + "created_at": "2026-09-20T15:21:09Z", + "updated_at": "2026-09-21T01:46:50Z", + "homepage": "https://github.com/Xubqpanda/JevRepo", + "default_branch": "main", + "default_sha": "50236bf2221bdc9b3f556ad4f6e067daf558c12b", + "readme": { + "path": "README.md", + "sha": "0e3f67c72801630bee7149c7e8157ea1370c37fa", + "size": 11730 + }, + "topics": [ + "agent-framework", + "agent-loop", + "ai-agent", + "decision-model", + "jev", + "llm", + "system-one", + "typescript", + "zero-dependency" + ], + "desc_hash": "ab62fbd74fcf" + }, + { + "id": "YuanKJing/Jev-as-Policy", + "description": "The highly anticipated open-source repository for JEV as Policy enables one-click setup of the simulation environment. Evaluations of Astra + JEV on benchmarks such as RoboTwin will also be released soon.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:43:07Z", + "created_at": "2026-09-21T01:43:06Z", + "updated_at": "2026-09-21T01:48:09Z", + "homepage": "", + "default_branch": "main", + "default_sha": "c2e1e17b6bb05fdbea1284d00730597cca161443", + "readme": { + "path": "README.md", + "sha": "2b83b5db78bd8d1bf4acdeb90fe8ab8e04b7948f", + "size": 225 + }, + "topics": [], + "desc_hash": "86f0b2c6ad77" + }, + { + "id": "Zafer-Liu/jev-demos", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": "HTML", + "size": 0, + "pushed_at": "2026-09-21T01:14:29Z", + "created_at": "2026-09-21T01:14:24Z", + "updated_at": "2026-09-21T01:14:33Z", + "homepage": null, + "default_branch": "main", + "default_sha": "c875343d3780a79426854e7795abf0e63f7c3dfa", + "readme": { + "path": "README.md", + "sha": "b9976762011fbb55fa71f734f2a0da2c2a4fae5e", + "size": 3656 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "aboisvert/jevvy", + "description": "Use jev model to augment csv files with inferred classification, scoring, or probability scores", + "stars": 0, + "forks": 0, + "license": "Apache-2.0", + "language": "Scala", + "size": 38, + "pushed_at": "2026-09-21T01:26:21Z", + "created_at": "2026-09-21T00:51:25Z", + "updated_at": "2026-09-21T01:23:09Z", + "homepage": "https://github.com/aboisvert/jevvy", + "default_branch": "main", + "default_sha": "14de871f75034a217d2d9aa2567e3ff5e1445a5c", + "readme": { + "path": "README.md", + "sha": "939119dd6c04d66d820f85a1ee2b59cb79962339", + "size": 4592 + }, + "topics": [ + "classification", + "csv-processing", + "graalvm-native-image", + "jev", + "probability", + "scala", + "scoring" + ], + "desc_hash": "77d47c2ad02b" + }, + { + "id": "andrest04/jev-lab", + "description": "Local lab for learning and testing TypeSafe Jev (System One)", + "stars": 0, + "forks": 0, + "license": null, + "language": "JavaScript", + "size": 0, + "pushed_at": "2026-09-21T01:24:20Z", + "created_at": "2026-09-21T01:23:03Z", + "updated_at": "2026-09-21T01:24:21Z", + "homepage": null, + "default_branch": "main", + "default_sha": "3ef5eb16c288bb41009d1c3174e7b0a6e1225b3e", + "readme": { + "path": "README.md", + "sha": "741995e66b02d6a11aa07c40620189ced9d3a3e8", + "size": 2063 + }, + "topics": [], + "desc_hash": "a838c0156f8c" + }, + { + "id": "andyrewlee/awesome-system-one", + "description": "Curated list of tools related to system one models", + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 9, + "pushed_at": "2026-09-20T07:36:47Z", + "created_at": "2026-09-19T23:20:48Z", + "updated_at": "2026-09-20T07:36:51Z", + "homepage": null, + "default_branch": "main", + "default_sha": "1345d8d237d33b5f3a43e99899741a2cb090df41", + "readme": { + "path": "README.md", + "sha": "5bf4d8b34e084d9861a80c4196abe4533276ac2d", + "size": 11919 + }, + "topics": [], + "desc_hash": "77976d967f10" + }, + { + "id": "ashafizullah/jev-triage", + "description": "Automated issue & PR triage for open-source maintainers, powered by Jev (TypeSafe AI).", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "TypeScript", + "size": 0, + "pushed_at": "2026-09-21T01:45:25Z", + "created_at": "2026-09-21T00:09:58Z", + "updated_at": "2026-09-21T01:45:28Z", + "homepage": null, + "default_branch": "main", + "default_sha": "6907ccc444a25c02f58f6f3dfaeec281bec76fa6", + "readme": { + "path": "README.md", + "sha": "96dcb0a587985b7c90bac63b25a28b625c1b3643", + "size": 16627 + }, + "topics": [], + "desc_hash": "9dead8b5a1bc" + }, + { + "id": "blanket11/jev-guide-ja", + "description": "jev\u306e\u5b66\u7fd2", + "stars": 0, + "forks": 0, + "license": null, + "language": "JavaScript", + "size": 0, + "pushed_at": "2026-09-21T01:37:48Z", + "created_at": "2026-09-21T01:15:10Z", + "updated_at": "2026-09-21T01:37:51Z", + "homepage": null, + "default_branch": "main", + "default_sha": "d844446d83867f0a93b72c74a1ded11ad130ee6a", + "readme": { + "path": "README.md", + "sha": "b6702287356c89aa2fe0e2735e8ac062897a1a19", + "size": 8518 + }, + "topics": [], + "desc_hash": "7006ad0d6b76" + }, + { + "id": "caohy1988/jev-guard-smoke", + "description": "Lab smoke proof for leepokai/jev-guard (Awesome-Jev #1). No API keys.", + "stars": 0, + "forks": 0, + "license": null, + "language": "Shell", + "size": 0, + "pushed_at": "2026-09-21T01:47:42Z", + "created_at": "2026-09-21T01:42:42Z", + "updated_at": "2026-09-21T01:42:52Z", + "homepage": null, + "default_branch": "docs/jev-guard-smoke-lab", + "default_sha": "8d3f629e1cb20ef55e2c379cca3379625f45d609", + "readme": null, + "topics": [], + "desc_hash": "3ad3b889d119" + }, + { + "id": "dinkarjuyal/jev-gepa", + "description": "Wiring a fast local NLI judge (Jev) into GEPA's reflective prompt optimization loop", + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:34:10Z", + "created_at": "2026-09-21T01:34:06Z", + "updated_at": "2026-09-21T01:34:14Z", + "homepage": null, + "default_branch": "main", + "default_sha": "3d52ba9b22397cb827b4f682db465ddf7d9aaffb", + "readme": { + "path": "README.md", + "sha": "655920a0144af6584d044638a77b83f3c81c1cde", + "size": 6297 + }, + "topics": [], + "desc_hash": "c4e769e2e075" + }, + { + "id": "early-effect/hexis", + "description": "ZIO / Scala 3 SDK for TypeSafe System One (Jev)", + "stars": 0, + "forks": 0, + "license": "Apache-2.0", + "language": "Scala", + "size": 0, + "pushed_at": "2026-09-21T01:54:48Z", + "created_at": "2026-09-21T01:50:01Z", + "updated_at": "2026-09-21T01:54:53Z", + "homepage": "https://www.earlyeffect.rocks/hexis/", + "default_branch": "main", + "default_sha": "c2d622c130baed8876f8f7fdca23e8d33c83deec", + "readme": { + "path": "README.md", + "sha": "4ffae17a7060ebc510508b4b5aab48b475fdcc96", + "size": 1468 + }, + "topics": [], + "desc_hash": "9c215f1aee74" + }, + { + "id": "elberacasa/omawish", + "description": "A System One for Omarchy: type what you want, and your desktop does it. Local, instant, 33M parameters, fine-tuned on one gaming GPU.", + "stars": 0, + "forks": 0, + "license": "NOASSERTION", + "language": "Rust", + "size": 0, + "pushed_at": "2026-09-21T01:48:04Z", + "created_at": "2026-09-21T01:34:23Z", + "updated_at": "2026-09-21T01:51:59Z", + "homepage": "https://github.com/elberacasa/omawish#readme", + "default_branch": "main", + "default_sha": "82c43e5425df4acf75459a9350115d0a849c5f0b", + "readme": { + "path": "README.md", + "sha": "bb2be38501638237b1b16ea70f1302b48b78861d", + "size": 17950 + }, + "topics": [ + "command-palette", + "hyprland", + "linux-desktop", + "local-first", + "omarchy", + "quickshell", + "rust", + "sentence-transformers", + "system-one" + ], + "desc_hash": "e53948cca959" + }, + { + "id": "endman100/research-Qwen3.8-JevLike", + "description": "71-label binary routing vs JSON Schema on Qwen3.8 NVFP4 / RTX 5090: measurements, raw evidence and ideal-parallel analysis", + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:31:38Z", + "created_at": "2026-09-21T01:18:14Z", + "updated_at": "2026-09-21T01:31:42Z", + "homepage": null, + "default_branch": "main", + "default_sha": "b406b17c3936e23ba588943cf05a15b8b27e0943", + "readme": { + "path": "README.md", + "sha": "7c908289d0e8a5ba36dbef64034ecf281dd2d89a", + "size": 5354 + }, + "topics": [], + "desc_hash": "90a2bc4aa36a" + }, + { + "id": "eteen12/jev-browser-automation", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:14:02Z", + "created_at": "2026-09-21T01:14:01Z", + "updated_at": "2026-09-21T01:14:02Z", + "homepage": null, + "default_branch": "main", + "default_sha": null, + "readme": null, + "topics": [], + "desc_hash": null + }, + { + "id": "fruitymcdoo/JevChat", + "description": "A chat interface built on Jev, TypeSafe's decision-only model: every word is a typed decision", + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:56:41Z", + "created_at": "2026-09-21T01:40:58Z", + "updated_at": "2026-09-21T01:56:45Z", + "homepage": null, + "default_branch": "main", + "default_sha": "6e2bc022cbbec86f055b57f6e67195642390434a", + "readme": { + "path": "README.md", + "sha": "abb03d83190aedc8055dda09aa1f79aa7dd2ef51", + "size": 11345 + }, + "topics": [], + "desc_hash": "dcde87688f6c" + }, + { + "id": "ismaelsoilet/jev-harness", + "description": null, + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:49:05Z", + "created_at": "2026-09-21T01:48:13Z", + "updated_at": "2026-09-21T01:49:08Z", + "homepage": null, + "default_branch": "main", + "default_sha": "9686bfd363c7f3bf818ea94c474e4df5e9e4e939", + "readme": { + "path": "README.md", + "sha": "7d6a0078ec5f306dec87b1508d1292dbdeda6f95", + "size": 12696 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "jacks3tr/Jev-Desktop", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:27:39Z", + "created_at": "2026-09-21T01:27:39Z", + "updated_at": "2026-09-21T01:27:39Z", + "homepage": null, + "default_branch": "main", + "default_sha": null, + "readme": null, + "topics": [], + "desc_hash": null + }, + { + "id": "javsanesq/jevlab", + "description": "A terminal workbench for learning, testing, and connecting TypeSafe Jev decisions", + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 548, + "pushed_at": "2026-09-21T00:58:07Z", + "created_at": "2026-09-21T00:09:01Z", + "updated_at": "2026-09-21T00:53:43Z", + "homepage": null, + "default_branch": "main", + "default_sha": "df7198de170086c2714132fdbd67061b4cfb764f", + "readme": { + "path": "README.md", + "sha": "87d810d6a4b2633840606f0c18dd2313feaaeff2", + "size": 30431 + }, + "topics": [], + "desc_hash": "65f3b669a7ca" + }, + { + "id": "joelakaufmann-lgtm/NRS-Navigator", + "description": "Local Nevada statute search and a reproducible evaluation of Jev-assisted ranking against keyword search.", + "stars": 0, + "forks": 0, + "license": "Apache-2.0", + "language": "HTML", + "size": 0, + "pushed_at": "2026-09-21T01:45:20Z", + "created_at": "2026-09-21T01:45:05Z", + "updated_at": "2026-09-21T01:45:26Z", + "homepage": null, + "default_branch": "main", + "default_sha": "6c9394e6c6cb78b4c0df6371a8b48eea5f6b47b9", + "readme": { + "path": "README.md", + "sha": "dbe9da1004bdce5116f5aa9cee46ead9b40e37bd", + "size": 17445 + }, + "topics": [], + "desc_hash": "5c3f19d9c202" + }, + { + "id": "k-srkw/jev-playground", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:38:24Z", + "created_at": "2026-09-21T01:25:57Z", + "updated_at": "2026-09-21T01:38:22Z", + "homepage": null, + "default_branch": "main", + "default_sha": "d935d01d70ab8c23a500f8006d57928124def021", + "readme": { + "path": "README.md", + "sha": "0b0945882b0c32c891aa354d9209ac5561ea9c19", + "size": 16 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "koteitan/laya-bot-det", + "description": "bot detector by laya for nostr", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": null, + "size": 1, + "pushed_at": "2026-09-21T00:57:03Z", + "created_at": "2026-09-21T00:57:02Z", + "updated_at": "2026-09-21T00:57:07Z", + "homepage": null, + "default_branch": "main", + "default_sha": "93120b7f808abe98c623ea4a7c9ebae16c809aa4", + "readme": null, + "topics": [], + "desc_hash": "c3f47e577eed" + }, + { + "id": "kzkhykw/jev-or-not", + "description": "Jev\u308b\uff1f\u30e2\u30c7\u30eb\u9078\u629e\u30d5\u30ed\u30fc\u30c1\u30e3\u30fc\u30c8", + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:13:53Z", + "created_at": "2026-09-21T01:04:37Z", + "updated_at": "2026-09-21T01:13:57Z", + "homepage": null, + "default_branch": "jev", + "default_sha": "916a83d68c9ae8ef4232c3ae86a1aee45c695a06", + "readme": { + "path": "README.md", + "sha": "7cbe3a508fab5775105262379189e6f0011722d2", + "size": 1120 + }, + "topics": [], + "desc_hash": "a3279deb950a" + }, + { + "id": "laidick/system-one-benchmark", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 14, + "pushed_at": "2026-09-19T23:48:40Z", + "created_at": "2026-09-19T23:48:35Z", + "updated_at": "2026-09-19T23:48:44Z", + "homepage": null, + "default_branch": "main", + "default_sha": "ccd4b79d1a516420450cd0811b86fd7aa528ef05", + "readme": { + "path": "README.md", + "sha": "5ffe72671473b1a0462e0dc7f05710e040eedc46", + "size": 10650 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "lezgoverci/jev-docs", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 410, + "pushed_at": "2026-09-21T00:04:09Z", + "created_at": "2026-09-21T00:04:01Z", + "updated_at": "2026-09-21T00:04:13Z", + "homepage": null, + "default_branch": "main", + "default_sha": "db9c9a85d63d2f0bae1d52856716a6a7f1b1d2e9", + "readme": { + "path": "README.md", + "sha": "f06bc61d36e9db7d4677fde855cc2bdd710eadf1", + "size": 14802 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "luckberonne/mini-jev", + "description": "Clasificador de comandos de shell de una sola pasada (solo lectura / reversible / destructivo), inspirado en Jev", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 54, + "pushed_at": "2026-09-21T00:01:26Z", + "created_at": "2026-09-21T00:01:21Z", + "updated_at": "2026-09-21T00:01:30Z", + "homepage": null, + "default_branch": "master", + "default_sha": "97d4b30b6f5de282c66809c4a34c5ea164454995", + "readme": { + "path": "README.md", + "sha": "795ab56ea245d37dde4c362999a6cc9a42c7d602", + "size": 2630 + }, + "topics": [], + "desc_hash": "06269d8dace1" + }, + { + "id": "marcosmartinez/jev-acento", + "description": "\u00bfJev entiende tu acento? Pre-registered audit of TypeSafe AI's Jev on Spanish \u2014 accuracy, calibration and token cost \u2014 plus a CLI to run the same comparison on your own labelled data.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 2285, + "pushed_at": "2026-09-21T01:12:19Z", + "created_at": "2026-09-20T23:23:21Z", + "updated_at": "2026-09-21T01:12:46Z", + "homepage": "https://github.com/marcosmartinez/jev-acento/releases/latest", + "default_branch": "main", + "default_sha": "7e007b4c2bd552471b56eff035f9a3df7593f9fe", + "readme": { + "path": "README.md", + "sha": "994943244b6ba1fdaf3420b2dcf5a9066f158509", + "size": 12639 + }, + "topics": [ + "benchmark", + "calibration", + "expected-calibration-error", + "jev", + "llm-evaluation", + "nlp", + "pre-registration", + "reproducible-research", + "spanish-nlp", + "typesafe-ai" + ], + "desc_hash": "1f8cc7248824" + }, + { + "id": "mednabouli/jev-ai-polymarket-copy-trading", + "description": "Automated Polymarket copy trading bot with MCP servers, Telegram alerts, and profitable wallet tracking. Zero API keys - uses Claude Code OAuth session auth.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 174, + "pushed_at": "2026-09-21T01:59:45Z", + "created_at": "2026-09-20T23:51:24Z", + "updated_at": "2026-09-21T01:59:48Z", + "homepage": null, + "default_branch": "main", + "default_sha": "aad77195fff5e1b4dafc9474a2f21239c824f33e", + "readme": { + "path": "README.md", + "sha": "7bb1a24b410018502e3263e43b1646b24008426e", + "size": 4581 + }, + "topics": [], + "desc_hash": "7082b3ecee24" + }, + { + "id": "olivdx/jev-mcp", + "description": "Jev-powered decision layer for coding agents. Analyze code and diffs, assess bugs, security, risk, and breaking changes, and return structured decisions for automated continue, fix, retry, or human-review workflows.", + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:42:36Z", + "created_at": "2026-09-21T01:42:35Z", + "updated_at": "2026-09-21T01:42:36Z", + "homepage": null, + "default_branch": "main", + "default_sha": null, + "readme": null, + "topics": [], + "desc_hash": "418b7804fdf5" + }, + { + "id": "pattoor/JEV-agent-opencv", + "description": "Juego simple para probar el modelo JEV con vision", + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:18:22Z", + "created_at": "2026-09-21T01:18:21Z", + "updated_at": "2026-09-21T01:18:22Z", + "homepage": null, + "default_branch": "main", + "default_sha": null, + "readme": null, + "topics": [], + "desc_hash": "9b5f38d0b031" + }, + { + "id": "peach-zhang/typesafe-go", + "description": "TypeSafe System One (Jev) \u7684 Go SDK \u2014 \u7c7b\u578b\u5316\u5224\u65ad\u4e0e\u6982\u7387,\u4ee3\u7801\u638c\u63a7\u5de5\u4f5c\u6d41", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Go", + "size": 0, + "pushed_at": "2026-09-21T01:42:10Z", + "created_at": "2026-09-21T01:34:16Z", + "updated_at": "2026-09-21T01:42:13Z", + "homepage": null, + "default_branch": "main", + "default_sha": "cd4b6bf028f73244440b898c5704b33bfa95f7af", + "readme": { + "path": "README.md", + "sha": "109ec34a82475c67df79d5ad2400d1314662d8e9", + "size": 6520 + }, + "topics": [], + "desc_hash": "9dd9a4e347ce" + }, + { + "id": "promptgtm-shared/clay-jev-people-ranker", + "description": "Agent Skill and Python workflow for Clay lead scoring, B2B prospect qualification, and people-search ranking with TypeSafe JEV.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:22:20Z", + "created_at": "2026-09-21T01:15:47Z", + "updated_at": "2026-09-21T01:22:23Z", + "homepage": null, + "default_branch": "main", + "default_sha": "2494b65218d4e4e940911db7c485710a3cff7050", + "readme": { + "path": "README.md", + "sha": "40402cc861cb84d1ace9110cabe24405505dd536", + "size": 12210 + }, + "topics": [ + "agent-skills", + "ai-agents", + "ai-lead-scoring", + "b2b-sales", + "claude-code", + "clay", + "clay-cli", + "cursor", + "grok", + "gtm-engineering", + "jev", + "lead-qualification", + "lead-scoring", + "people-search", + "prospecting", + "python", + "revops", + "sales-automation", + "typesafe-ai", + "typesafe-jev" + ], + "desc_hash": "1d3275407722" + }, + { + "id": "sahasrarjn/system-one", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 62, + "pushed_at": "2026-09-20T01:51:11Z", + "created_at": "2026-09-19T06:07:11Z", + "updated_at": "2026-09-20T01:51:15Z", + "homepage": null, + "default_branch": "main", + "default_sha": "bb2383834d198ca00ffb498c063ba1b455e1ff87", + "readme": { + "path": "README.md", + "sha": "b71d9004231903688f450731977d5a482ff777ae", + "size": 8044 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "sarathi-aiml/jevsql", + "description": "Text-to-SQL where the model never writes SQL \u2014 typed, calibrated decisions (TypeSafe Jev) + code-assembled queries", + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:50:41Z", + "created_at": "2026-09-21T01:28:29Z", + "updated_at": "2026-09-21T01:50:45Z", + "homepage": null, + "default_branch": "main", + "default_sha": "ba46b8b400a1e13dfa468bbfce2f370cf9d01972", + "readme": { + "path": "README.md", + "sha": "9c05b06bf11aac2ed94e953eb71fcfd415e56c5c", + "size": 10324 + }, + "topics": [], + "desc_hash": "b5cb26f386df" + }, + { + "id": "theosunny/jev_stock", + "description": null, + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 54, + "pushed_at": "2026-09-21T01:48:42Z", + "created_at": "2026-09-20T09:11:00Z", + "updated_at": "2026-09-21T01:48:45Z", + "homepage": null, + "default_branch": "main", + "default_sha": "0c0e3fbbb7861432e70a724036e4b26e12f03e42", + "readme": { + "path": "README.md", + "sha": "063036893cb5832a65843187a7373b734535af9a", + "size": 3924 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "twilwa/pi-typesafe", + "description": "Pi coding-agent extension built on the TypeSafe AI System One API (Jev)", + "stars": 0, + "forks": 0, + "license": null, + "language": "TypeScript", + "size": 94, + "pushed_at": "2026-09-21T01:43:14Z", + "created_at": "2026-09-17T00:18:58Z", + "updated_at": "2026-09-21T01:43:18Z", + "homepage": null, + "default_branch": "main", + "default_sha": "3dc40b969dc3c8f90e313622c3e69a91b63b7cc6", + "readme": { + "path": "README.md", + "sha": "195563911db2f04799efe12a707a66e695eb46c3", + "size": 20418 + }, + "topics": [], + "desc_hash": "d86a28e2715b" + }, + { + "id": "willgriffin/pi-fusion-matrix", + "description": "Multi-model deliberation for the pi coding agent: named fusions, per-slot fallback and routing, version-free aliases, conservative decision backends", + "stars": 0, + "forks": 0, + "license": null, + "language": "JavaScript", + "size": 592, + "pushed_at": "2026-09-21T01:46:54Z", + "created_at": "2026-09-18T18:53:12Z", + "updated_at": "2026-09-21T01:46:58Z", + "homepage": null, + "default_branch": "main", + "default_sha": "f5d2a1b89448bd2dd84ae563a8f1d5b613e11425", + "readme": { + "path": "README.md", + "sha": "dfb8b07ad2c06d212db224e664601ec9bad42958", + "size": 49676 + }, + "topics": [], + "desc_hash": "0fadc538568e" + }, + { + "id": "alexwestco/llm-to-jev", + "description": "Convert LLM prompts to Jev prompts", + "stars": 3, + "forks": 0, + "license": "MIT", + "language": "JavaScript", + "size": 29, + "pushed_at": "2026-09-20T12:06:01Z", + "created_at": "2026-09-19T15:20:36Z", + "updated_at": "2026-09-20T17:23:58Z", + "homepage": null, + "default_branch": "main", + "default_sha": "234058ab372d7754833c6279601755f8fda55d98", + "readme": { + "path": "README.md", + "sha": "43cd94fba7527d78635b148628485c0cd1b1ad66", + "size": 3637 + }, + "topics": [], + "desc_hash": "9f521cf7cc20" + } +] \ No newline at end of file diff --git a/research/archive/hourly/2026-09-21T01/novel_high_this_run.json b/research/archive/hourly/2026-09-21T01/novel_high_this_run.json new file mode 100644 index 0000000..36a72f4 --- /dev/null +++ b/research/archive/hourly/2026-09-21T01/novel_high_this_run.json @@ -0,0 +1,1746 @@ +[ + { + "id": "brainstormity/Jev-X-Sentiment-Analysis", + "description": null, + "stars": 136, + "forks": 20, + "license": null, + "language": "Python", + "size": 467, + "pushed_at": "2026-09-19T22:51:11Z", + "created_at": "2026-09-19T22:51:07Z", + "updated_at": "2026-09-21T01:57:49Z", + "homepage": null, + "default_branch": "main", + "default_sha": "5c932f941a92348781e8e5b471a9ddb6af980253", + "readme": { + "path": "README.md", + "sha": "bf4134b44cda73599f005e9408391353a2ed9437", + "size": 7370 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "ikermoel/open-alternative-jev", + "description": "Open-source alternative to TypeSafe's Jev: a System One style model layer that gives typed, calibrated decisions from any open-weights LLM in one forward pass (HF + vLLM), with honest benchmarks", + "stars": 36, + "forks": 8, + "license": "Apache-2.0", + "language": "Python", + "size": 2025, + "pushed_at": "2026-09-20T03:14:39Z", + "created_at": "2026-09-18T05:20:56Z", + "updated_at": "2026-09-20T23:48:09Z", + "homepage": "https://huggingface.co/spaces/IkerMoel/open-alternative-jev", + "default_branch": "main", + "default_sha": "6ad87d7ce2f4ef472acb9253419134a2643a57d2", + "readme": { + "path": "README.md", + "sha": "d4653d9d1b95a0d584ea1fb9c7e11426463d7fd6", + "size": 11764 + }, + "topics": [ + "calibration", + "classification", + "jev", + "jev-alternative", + "llm", + "logprobs", + "open-jev", + "open-source-jev", + "open-weights", + "qwen", + "structured-decisions", + "structured-output", + "system-one", + "system-one-model", + "transformers", + "typed-decisions", + "typesafe", + "typesafe-jev", + "vllm" + ], + "desc_hash": "3aeedb10b4e9" + }, + { + "id": "okinaaudio/live-jev", + "description": "Control Ableton Live with one short sentence (Japanese / English). Summon with \u2318\u21e7Space, type or dictate, done.", + "stars": 36, + "forks": 2, + "license": "MIT", + "language": "Python", + "size": 522, + "pushed_at": "2026-09-21T01:46:35Z", + "created_at": "2026-09-19T03:15:09Z", + "updated_at": "2026-09-21T01:46:40Z", + "homepage": null, + "default_branch": "main", + "default_sha": "2446eb777ad9f59f77b96ee5b081ee8c2812e0a3", + "readme": { + "path": "README.md", + "sha": "52e7d45c799f9b4d20053d2c7e28b8caac0e1b07", + "size": 12968 + }, + "topics": [ + "ableton-live", + "macos", + "music-production", + "natural-language" + ], + "desc_hash": "73f6f5c579e5" + }, + { + "id": "heyjunpenn/awesome-jev", + "description": "A verified, community-maintained catalog of 503 open-source projects built with Jev.", + "stars": 32, + "forks": 2, + "license": "MIT", + "language": "TypeScript", + "size": 3326, + "pushed_at": "2026-09-20T10:08:27Z", + "created_at": "2026-09-19T05:14:36Z", + "updated_at": "2026-09-21T01:47:12Z", + "homepage": "https://jevbest.com/", + "default_branch": "main", + "default_sha": "8ecdef6a6fcb3acc0c498fe8a87855e5bcd21c0f", + "readme": { + "path": "README.md", + "sha": "0b24293bc84ff0cc828b67c7d5049b74d0d085ed", + "size": 104955 + }, + "topics": [], + "desc_hash": "eed807328057" + }, + { + "id": "NanmiCoder/jev-arena", + "description": "Jev \u6a21\u578b\u4ecb\u7ecd\u4e0e\u5b9e\u6d4b\uff1a\u901a\u8fc7 Choice / Score / Noul \u5c06\u81ea\u7136\u8bed\u8a00\u8f6c\u4e3a\u5e26\u7c7b\u578b\u7684\u5224\u65ad\u4e0e\u6982\u7387\uff0c\u7528\u4e8e\u5206\u7c7b\u3001\u8bc4\u5206\u548c\u8def\u7531\uff1b\u652f\u6301\u4e0e DeepSeek \u7b49\u6a21\u578b\u5bf9\u6bd4\u8bc4\u8bba\u6253\u6807\u3001\u901f\u5ea6\u4e0e\u7ed3\u679c\uff0c\u542b CSV/Excel \u5bfc\u5165\u3001\u539f\u901f\u56de\u653e\u4e0e\u79bb\u7ebf\u62a5\u544a\u3002", + "stars": 31, + "forks": 3, + "license": "MIT", + "language": "JavaScript", + "size": 18427, + "pushed_at": "2026-09-20T15:00:39Z", + "created_at": "2026-09-19T11:01:19Z", + "updated_at": "2026-09-21T01:37:34Z", + "homepage": "https://nanmicoder.github.io/jev-arena/", + "default_branch": "main", + "default_sha": "2ca160cc4aa9ac72a4341e2ac5903258e8c69c84", + "readme": { + "path": "README.md", + "sha": "4eb7f2dec20a2ecdf0d74a6254f311ee8bff19a7", + "size": 2878 + }, + "topics": [], + "desc_hash": "e2732659dccb" + }, + { + "id": "yzfly/awesome-jev-zh", + "description": "Jev / TypeSafe System One \u4e2d\u6587\u7cbe\u9009\u5217\u8868\uff1a\u5b98\u65b9\u8d44\u6599\u3001SDK\u3001\u7206\u6b3e\u5e94\u7528\u3001Agent \u5de5\u5177\u3001\u5f00\u6e90\u590d\u73b0\u4e0e\u72ec\u7acb\u8bc4\u6d4b\uff0c\u9644\u4e2d\u6587\u4e0a\u624b\u6307\u5357\uff0c\u6bcf\u65e5\u81ea\u52a8\u6536\u5f55 GitHub \u70ed\u95e8\u9879\u76ee\u3002", + "stars": 31, + "forks": 9, + "license": "CC0-1.0", + "language": "HTML", + "size": 515, + "pushed_at": "2026-09-21T01:20:52Z", + "created_at": "2026-09-18T06:33:02Z", + "updated_at": "2026-09-21T01:41:00Z", + "homepage": "https://code.jiangshu.ai/awesome-jev-zh/", + "default_branch": "main", + "default_sha": "7cce63fedfc507993f311ced0f270e1a2739695a", + "readme": { + "path": "README.md", + "sha": "f8c549471e6ae4bbc3525408fda7c7e1d1feb626", + "size": 106685 + }, + "topics": [ + "ai-agent", + "awesome", + "awesome-list", + "chinese", + "jev", + "llm", + "structured-output", + "system-one", + "typesafe", + "typesafe-ai" + ], + "desc_hash": "1ee1f7288a28" + }, + { + "id": "openroboto-ai/jev-robot-control", + "description": null, + "stars": 27, + "forks": 0, + "license": "NOASSERTION", + "language": "Python", + "size": 18965, + "pushed_at": "2026-09-19T13:39:07Z", + "created_at": "2026-09-19T13:35:41Z", + "updated_at": "2026-09-20T20:15:14Z", + "homepage": null, + "default_branch": "main", + "default_sha": "7a4ed8b72c3c17d7aa790678ed9660df67c10dd3", + "readme": { + "path": "README.md", + "sha": "0e749c38197e78c46dacbb1f61910ef2b244a823", + "size": 8306 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "win4r/jev-skill-suggester", + "description": "\u7528 TypeSafe Jev \u63a8\u8350\u5df2\u5b89\u88c5 Skill / Bounded installed-skill recommendations with TypeSafe Jev. Python CLI, Codex skill, bilingual docs and live examples.", + "stars": 27, + "forks": 2, + "license": "MIT", + "language": "Python", + "size": 47, + "pushed_at": "2026-09-19T15:31:40Z", + "created_at": "2026-09-19T15:31:35Z", + "updated_at": "2026-09-21T00:32:21Z", + "homepage": null, + "default_branch": "main", + "default_sha": "05fbd7ce9ec74a2f193c09276a8d4d077d9d5e5c", + "readme": { + "path": "README.md", + "sha": "2a03bec83f3c9915d7c8691523a7a489c65fee7e", + "size": 11404 + }, + "topics": [], + "desc_hash": "a146a470c049" + }, + { + "id": "PyModel/typesafe-mcp", + "description": null, + "stars": 23, + "forks": 2, + "license": "MIT", + "language": "Go", + "size": 1123, + "pushed_at": "2026-09-21T01:48:45Z", + "created_at": "2026-09-19T18:43:24Z", + "updated_at": "2026-09-21T01:48:36Z", + "homepage": null, + "default_branch": "main", + "default_sha": "cf01808bf27485a017cc6eebead6384556dab552", + "readme": { + "path": "README.md", + "sha": "c7c36760aa1cfb1c4db3bafdc867b4bafd1dd241", + "size": 15857 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "AkashPriyadarshii/jev-seo", + "description": "100% free \u20b90 agent-first SEO & GEO CLI suite and MCP server in Rust replacing Semrush and OpenSEO via DuckDuckGo and TypeSafe Jev System One", + "stars": 21, + "forks": 1, + "license": "MIT", + "language": "Rust", + "size": 102, + "pushed_at": "2026-09-21T01:51:33Z", + "created_at": "2026-09-18T13:02:32Z", + "updated_at": "2026-09-21T01:50:37Z", + "homepage": "https://jevseo.vercel.app", + "default_branch": "master", + "default_sha": "f8cb7c55c356cbc12929cf548ae75ea568f03861", + "readme": { + "path": "README.md", + "sha": "e3290fd15adde74b7081fca7503e65e228306419", + "size": 11529 + }, + "topics": [ + "ahrefs-alternative", + "claude-code", + "cli", + "decision-oracle", + "duckduckgo", + "generative-engine-optimization", + "geo", + "jev", + "mcp", + "mcp-server", + "rank-tracker", + "rust", + "search-engine-optimization", + "semrush-alternative", + "seo", + "seo-audit", + "seo-tools", + "serp", + "typesafe", + "typesafe-ai" + ], + "desc_hash": "3ec093c905b3" + }, + { + "id": "devtooligan/jevscan-evm", + "description": null, + "stars": 21, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 575, + "pushed_at": "2026-09-19T20:15:42Z", + "created_at": "2026-09-19T00:25:35Z", + "updated_at": "2026-09-20T22:02:31Z", + "homepage": null, + "default_branch": "main", + "default_sha": "2ecbe7cd93a92adb4464821d3ca7b0828c0c1b3c", + "readme": { + "path": "README.md", + "sha": "075a01007b92aec41fd2a4ff52d1c69d07840028", + "size": 8763 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "mizzlelover/jev-hub", + "description": "JEV HUB \u00b7 X \u4e0a\u5173\u4e8e TypeSafe AI\u300c\u7cfb\u7edf\u4e00\u6a21\u578b\u300dJev \u7684\u957f\u6587\u4e0e\u6f14\u793a\u89c6\u9891\u805a\u5408\uff08\u4fdd\u7559\u539f\u94fe\u4e0e\u4f5c\u8005\uff09\uff5c \u8c01\u662f\u4e13\u5bb6 \u51fa\u54c1", + "stars": 21, + "forks": 3, + "license": "NOASSERTION", + "language": "CSS", + "size": 189, + "pushed_at": "2026-09-19T02:54:57Z", + "created_at": "2026-09-19T02:32:02Z", + "updated_at": "2026-09-20T07:21:23Z", + "homepage": "https://mizzlelover.github.io/jev-hub/", + "default_branch": "main", + "default_sha": "4262d71d2d06a910349d709b4d4958ce34c5ed98", + "readme": { + "path": "README.md", + "sha": "04530e25b75136c90c0e65569dc13fb07a55aab1", + "size": 55472 + }, + "topics": [ + "ai", + "awesome-list", + "jev", + "typesafe", + "x-twitter" + ], + "desc_hash": "a5b7621eaac0" + }, + { + "id": "sgoedecke/system-one", + "description": "Batched single-token choice inference for open language models, compatible with TypeSafe", + "stars": 20, + "forks": 3, + "license": null, + "language": "Python", + "size": 125589, + "pushed_at": "2026-09-18T03:22:07Z", + "created_at": "2026-09-17T12:13:54Z", + "updated_at": "2026-09-20T21:23:45Z", + "homepage": null, + "default_branch": "main", + "default_sha": "ebde2a2db7067b920dfe51e9ce785613e66613d5", + "readme": { + "path": "README.md", + "sha": "d331b567e2c37146a6194004ee494d5f9bbc345d", + "size": 3331 + }, + "topics": [], + "desc_hash": "88c6ac90a44e" + }, + { + "id": "zszz3/Pi-Jev-Guide", + "description": null, + "stars": 19, + "forks": 0, + "license": "MIT", + "language": "TypeScript", + "size": 82, + "pushed_at": "2026-09-20T03:41:56Z", + "created_at": "2026-09-19T09:22:39Z", + "updated_at": "2026-09-20T08:42:41Z", + "homepage": null, + "default_branch": "main", + "default_sha": "f187d464062bfd6fb41ae690c15ad1af974cbfb6", + "readme": { + "path": "README.md", + "sha": "45aaa093f82bf3cd81b93451e7b06d9bbbfd06d1", + "size": 14302 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "mithalouni/system-one-open", + "description": "Open replica of TypeSafe's Jev: typed calibrated decisions in one forward pass, on Gemma 4 E2B / Gemma 3 270M (Modal)", + "stars": 18, + "forks": 4, + "license": "NOASSERTION", + "language": "Python", + "size": 13272, + "pushed_at": "2026-09-17T07:06:55Z", + "created_at": "2026-09-17T05:56:58Z", + "updated_at": "2026-09-20T18:12:14Z", + "homepage": null, + "default_branch": "main", + "default_sha": "77f1f7cccf8aa752e0ed7edcc8d2094bac707bdc", + "readme": { + "path": "README.md", + "sha": "535f33028a685b909fbc287409697be55da6756a", + "size": 5511 + }, + "topics": [], + "desc_hash": "8755f1023081" + }, + { + "id": "davila7/jev-explained", + "description": "Jev Explained", + "stars": 14, + "forks": 1, + "license": "MIT", + "language": "TypeScript", + "size": 1180, + "pushed_at": "2026-09-20T16:53:26Z", + "created_at": "2026-09-19T11:46:20Z", + "updated_at": "2026-09-21T00:04:33Z", + "homepage": "https://jev-explained-drab.vercel.app", + "default_branch": "main", + "default_sha": "5cbe35e04609112be77b1bd447bd79b3bde7980b", + "readme": { + "path": "README.md", + "sha": "7c69d941dc821e2df3a37e3348357f04b0b2adb8", + "size": 5548 + }, + "topics": [], + "desc_hash": "a20ee63f34c9" + }, + { + "id": "ckaraca/awesome-jev", + "description": "A curated list of tools, integrations, and experiments built on Jev, TypeSafe AI's System One model for fast, typed decisions.", + "stars": 7, + "forks": 1, + "license": "CC0-1.0", + "language": "Python", + "size": 47, + "pushed_at": "2026-09-20T13:48:37Z", + "created_at": "2026-09-19T00:52:55Z", + "updated_at": "2026-09-20T13:48:41Z", + "homepage": "https://docs.typesafe.ai/", + "default_branch": "main", + "default_sha": "2207d19d91fa80d54b4de3f86295437605be58e8", + "readme": { + "path": "README.md", + "sha": "3e58af61ccebd902207cc784eee84b8318d027d1", + "size": 20131 + }, + "topics": [ + "ai-agents", + "awesome", + "awesome-list", + "computer-use", + "jev", + "llm", + "system-one", + "typesafe" + ], + "desc_hash": "70b1ba2dc046" + }, + { + "id": "arunav25/jev-mcp", + "description": "Connect JEV to MCP clients and compare its judgments against general-purpose LLMs using shared datasets and measurable accuracy.", + "stars": 5, + "forks": 0, + "license": "MIT", + "language": "JavaScript", + "size": 69, + "pushed_at": "2026-09-17T16:05:17Z", + "created_at": "2026-09-17T15:19:03Z", + "updated_at": "2026-09-18T16:03:38Z", + "homepage": null, + "default_branch": "main", + "default_sha": "036d32433c0d234d37ef11a6a44619b69444556c", + "readme": { + "path": "README.md", + "sha": "a858bf2227da3012651457b31404f70f08fb807f", + "size": 10859 + }, + "topics": [], + "desc_hash": "adbf6ad85063" + }, + { + "id": "cobusgreyling/Jev", + "description": "Unofficial TypeSafe Jev showcase \u2014 System One decisions, not chat.", + "stars": 3, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 347, + "pushed_at": "2026-09-20T07:32:08Z", + "created_at": "2026-09-19T10:35:48Z", + "updated_at": "2026-09-20T07:32:11Z", + "homepage": "https://docs.typesafe.ai/", + "default_branch": "main", + "default_sha": "c636087edf4100a097a456e559c4b2fa769b80e8", + "readme": { + "path": "README.md", + "sha": "9de601117711b440fc64776052511a2dbc099240", + "size": 9262 + }, + "topics": [ + "calibrated-decisions", + "jev", + "showcase", + "system-one", + "typesafe" + ], + "desc_hash": "086eb9b31d04" + }, + { + "id": "nrdz-labs/fast-jev-opencode", + "description": "Jev-scored context pruning for OpenCode: drops stale tool calls and truncates bulky results on the outgoing request \u2014 fail-open, cache-backed, configurable live. Port of fast-jev-compaction to the V2 context hook.", + "stars": 3, + "forks": 0, + "license": "MIT", + "language": "TypeScript", + "size": 140, + "pushed_at": "2026-09-19T06:25:17Z", + "created_at": "2026-09-19T03:51:47Z", + "updated_at": "2026-09-20T22:59:18Z", + "homepage": null, + "default_branch": "main", + "default_sha": "4c4de4dcffa0201502fcb682c80b807cfd248140", + "readme": { + "path": "README.md", + "sha": "1a733604c7fb0757ae2e72e15c1829ef22deb366", + "size": 3814 + }, + "topics": [ + "compaction", + "context", + "jev", + "opencode", + "opencode-plugin", + "typesafe" + ], + "desc_hash": "50253fd57e4c" + }, + { + "id": "liao96312/jev-arena-nanojev", + "description": "\u5b8c\u5168\u672c\u5730\u7684 NanoJev \u7f51\u683c\u51b3\u7b56\u6e38\u620f\u5b9e\u9a8c\u573a\uff0c\u652f\u6301\u4e2d\u6587 Pygame\u3001\u591a\u5173\u5361\u4e0e GTX 1660S \u8bad\u7ec3", + "stars": 2, + "forks": 0, + "license": null, + "language": "Python", + "size": 13967, + "pushed_at": "2026-09-21T01:46:57Z", + "created_at": "2026-09-19T02:23:40Z", + "updated_at": "2026-09-21T01:47:01Z", + "homepage": null, + "default_branch": "main", + "default_sha": "a4d2ec811fc48d067a3c68ee9e976a2ee7d9a488", + "readme": { + "path": "README.md", + "sha": "a5887c4c76ff032ba3e5b7339bc12fc31e719f05", + "size": 13199 + }, + "topics": [], + "desc_hash": "0ceaea47dd48" + }, + { + "id": "Krug2/JevLM-Open", + "description": null, + "stars": 1, + "forks": 2, + "license": null, + "language": "Python", + "size": 10170, + "pushed_at": "2026-09-20T21:17:19Z", + "created_at": "2026-09-20T18:35:30Z", + "updated_at": "2026-09-21T00:44:46Z", + "homepage": null, + "default_branch": "main", + "default_sha": "84272bdabbea7ed76548b4fdd7d623a9373d01de", + "readme": null, + "topics": [], + "desc_hash": null + }, + { + "id": "Rizzo-AI-Academy/rizzo-flow", + "description": "The open, local take on Jev: typed decisions from an LLM, without generating a single token", + "stars": 1, + "forks": 0, + "license": "Apache-2.0", + "language": "Python", + "size": 1767, + "pushed_at": "2026-09-21T01:55:02Z", + "created_at": "2026-09-21T00:00:02Z", + "updated_at": "2026-09-21T01:55:06Z", + "homepage": "", + "default_branch": "main", + "default_sha": "d97ef676a40e2ad47541835f8676b4be60331911", + "readme": { + "path": "README.md", + "sha": "529432bebd5f6692f1ccffe6bbb0176dbc446aa9", + "size": 23362 + }, + "topics": [], + "desc_hash": "d006ffab3676" + }, + { + "id": "clouatre-labs/decisions-judge-mcp", + "description": "MCP server exposing an LLM judge (typed decisions: yes/no probability, choice, score) for coding agents", + "stars": 1, + "forks": 0, + "license": "Apache-2.0", + "language": "JavaScript", + "size": 114, + "pushed_at": "2026-09-21T01:49:37Z", + "created_at": "2026-09-20T22:01:23Z", + "updated_at": "2026-09-21T01:49:39Z", + "homepage": null, + "default_branch": "main", + "default_sha": "b6fff8ee714cfbbe94c6d9fc09dadfefd57a6676", + "readme": { + "path": "README.md", + "sha": "cd771f8872c011937bd41efdbf275669cfc0e6e8", + "size": 3383 + }, + "topics": [ + "ai", + "cli", + "decisions", + "judge", + "llm", + "mcp", + "mcp-server", + "model-context-protocol" + ], + "desc_hash": "79d7cb28cf57" + }, + { + "id": "emirbartu/opencode-system-one", + "description": "Opencode plugin using Jev (system one model) as part of software development process. Not affiliated with Opencode team.", + "stars": 1, + "forks": 0, + "license": null, + "language": "TypeScript", + "size": 70, + "pushed_at": "2026-09-19T00:28:05Z", + "created_at": "2026-09-19T00:28:02Z", + "updated_at": "2026-09-19T01:30:22Z", + "homepage": "", + "default_branch": "main", + "default_sha": "e51f5fde5ed235004970c3a90fd98033c3dcfdb5", + "readme": { + "path": "README.md", + "sha": "3c26ede85bd1644eca0787085d68af0286f085f2", + "size": 2910 + }, + "topics": [ + "jev", + "opencode-plugin", + "opencode-plugins" + ], + "desc_hash": "791b87501e61" + }, + { + "id": "kotoba-lang/typed-decisions", + "description": "Jev-shaped typed-decision model (state + Choice/Score/Noul questions -> calibrated probabilities, one pass) on ModernBERT / DeBERTa / LLaDA-MoE, with measured latency, accuracy, calibration and training cost", + "stars": 1, + "forks": 0, + "license": "NOASSERTION", + "language": "Python", + "size": 14217, + "pushed_at": "2026-09-20T02:58:24Z", + "created_at": "2026-09-18T07:03:52Z", + "updated_at": "2026-09-20T02:57:55Z", + "homepage": null, + "default_branch": "main", + "default_sha": "10d7834d3b99041f890db4615fb38ef95ced50cc", + "readme": { + "path": "README.md", + "sha": "4d6bbf4c4e446270dea99bb8b9d70bf2df628f09", + "size": 55530 + }, + "topics": [], + "desc_hash": "8588e6be9350" + }, + { + "id": "mallahyari/system-one-benchmark", + "description": null, + "stars": 1, + "forks": 0, + "license": null, + "language": "Python", + "size": 48, + "pushed_at": "2026-09-19T15:55:09Z", + "created_at": "2026-09-19T03:35:00Z", + "updated_at": "2026-09-19T15:55:12Z", + "homepage": null, + "default_branch": "main", + "default_sha": "46c232b887d41b4b560a1eac74972eade3b1fddc", + "readme": { + "path": "README.md", + "sha": "b4b8e76713c89ee429923a13c25577dd5c53f8cf", + "size": 6836 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "sable-inc/jev-linter-action", + "description": "Configurable semantic CI checks for repository files using TypeSafe Jev", + "stars": 1, + "forks": 0, + "license": "MIT", + "language": "JavaScript", + "size": 64, + "pushed_at": "2026-09-21T01:43:28Z", + "created_at": "2026-09-21T00:02:01Z", + "updated_at": "2026-09-21T01:48:23Z", + "homepage": null, + "default_branch": "main", + "default_sha": "1f6ba701fe72a70c2ec070c8eac3dff836a0d704", + "readme": { + "path": "README.md", + "sha": "b1ac97554b8269cacca8c0989cf402aa235e593d", + "size": 4791 + }, + "topics": [], + "desc_hash": "84b832b28033" + }, + { + "id": "shirenchuang/awsomejev", + "description": "Awesome Jev\uff1aJev \u5f00\u6e90\u751f\u6001\u5bfc\u822a", + "stars": 1, + "forks": 0, + "license": "MIT", + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:34:16Z", + "created_at": "2026-09-21T01:25:12Z", + "updated_at": "2026-09-21T01:44:17Z", + "homepage": null, + "default_branch": "main", + "default_sha": "a994099f0a32d027b7229fedf560eaf7a2bb3c46", + "readme": { + "path": "README.md", + "sha": "96a2075be5bb611ac82986ab611cea0806749c16", + "size": 52785 + }, + "topics": [], + "desc_hash": "c4be2e0491a3" + }, + { + "id": "1816586742-stack/jev-craft", + "description": "\u8ba9 Agent \u957f\u51fa\u300c\u64cd\u4f5c\u6746\u300d\uff1a\u7528 System One \u6a21\u578b\uff08Jev\uff09\u627f\u62c5\u9ad8\u9891\u5224\u65ad\u3001\u591a\u6a21\u6001\u6a21\u578b\u5f53\u773c\u775b\u3002\u7ed9\u601d\u8def + \u53ef\u8dd1\u7684\u53c2\u8003\u5b9e\u73b0\uff0875 \u6761\u79bb\u7ebf\u65ad\u8a00\uff0c\u96f6\u4f9d\u8d56\u96f6 key\uff09", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "JavaScript", + "size": 0, + "pushed_at": "2026-09-21T01:20:50Z", + "created_at": "2026-09-21T01:11:38Z", + "updated_at": "2026-09-21T01:20:54Z", + "homepage": null, + "default_branch": "main", + "default_sha": "ad4926f663b8c382f6db7e889288747ccf68d119", + "readme": { + "path": "README.md", + "sha": "b614d4f0b3133685b56d87d2eab4570abdedb92e", + "size": 13263 + }, + "topics": [ + "agent", + "ai-agent", + "decision-making", + "deepseek", + "dsh", + "jev", + "llm", + "minecraft", + "system-one", + "typesafe" + ], + "desc_hash": "5184161785c9" + }, + { + "id": "DolphinMiner/jev-rss", + "description": "A local-first RSS reader with Jev-powered semantic screening. Follow what matters, inspect every judgment, and keep control of your reading. English / \u7b80\u4f53\u4e2d\u6587.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "TypeScript", + "size": 0, + "pushed_at": "2026-09-21T01:03:30Z", + "created_at": "2026-09-21T01:02:51Z", + "updated_at": "2026-09-21T01:03:57Z", + "homepage": "", + "default_branch": "main", + "default_sha": "4a2dbb770faa3a9cac80d7b1f27d181f25b5bb91", + "readme": { + "path": "README.md", + "sha": "caaf2f1ebb6a62052ee95d268c16fb568aa58838", + "size": 8265 + }, + "topics": [ + "ai", + "jev", + "local-first", + "react", + "rss", + "rss-reader", + "rsshub", + "self-hosted", + "semantic-filtering", + "typescript" + ], + "desc_hash": "61001e76fee3" + }, + { + "id": "Dreydrey9000/jev-relay", + "description": "Guarded local-first decision advice for Claude Code, Codex and Hermes/Jax, with explicit Jev checks and review fallbacks.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 5213, + "pushed_at": "2026-09-21T00:59:23Z", + "created_at": "2026-09-21T00:47:38Z", + "updated_at": "2026-09-21T00:59:27Z", + "homepage": null, + "default_branch": "main", + "default_sha": "d52ffe1fc023ff607458c1aee1eba05d1f628f1e", + "readme": { + "path": "README.md", + "sha": "2039013eb53d6098a6c2dd9ecb2c46b52c37462a", + "size": 6375 + }, + "topics": [], + "desc_hash": "dff46c81ca68" + }, + { + "id": "Eric-Zhou-0302/jev-A-share-trader", + "description": "A Jev-powered technical analysis workspace for China A-shares, supporting AKShare/Tushare, market scans, and Buy/Hold/Sell assessments with time horizons and traceable evidence.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 217, + "pushed_at": "2026-09-21T01:43:35Z", + "created_at": "2026-09-20T04:53:28Z", + "updated_at": "2026-09-21T01:43:39Z", + "homepage": null, + "default_branch": "main", + "default_sha": "a7ad82306fbe8e0a2ef92a54f236684c7e4139d5", + "readme": { + "path": "README.md", + "sha": "9c1cd8d69c85c0e71c7159dfdcaefaab5ce4ed7b", + "size": 9451 + }, + "topics": [ + "a-shares", + "akshare", + "fastapi", + "jev", + "python", + "react", + "stock-analysis", + "system-one", + "technical-analysis", + "tushare", + "typesafe-ai" + ], + "desc_hash": "09fad6bc411d" + }, + { + "id": "JTech-CO/Jev-Simulink-Supervisor", + "description": "Jev-Simulink Supervisor", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "MATLAB", + "size": 0, + "pushed_at": "2026-09-21T01:54:08Z", + "created_at": "2026-09-21T01:15:44Z", + "updated_at": "2026-09-21T01:54:11Z", + "homepage": null, + "default_branch": "main", + "default_sha": "76bc14a89694a2eb275239a96e8df6044c18d6d2", + "readme": { + "path": "README.md", + "sha": "61f40464fa61b7b108e0a7dd5b48257e5dca70c3", + "size": 3211 + }, + "topics": [], + "desc_hash": "9e348cc4366c" + }, + { + "id": "Kwwwww74/OpenJev", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:31:34Z", + "created_at": "2026-09-21T01:31:33Z", + "updated_at": "2026-09-21T01:31:37Z", + "homepage": null, + "default_branch": "main", + "default_sha": "423875b6088bd2d9b4e639777e2bdafdd60c760e", + "readme": { + "path": "README.md", + "sha": "8f5ec58f92ea2ad2dddf9ad9e03e3153e38dc4a7", + "size": 9 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "RuipuCui/jev-harness", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T00:02:26Z", + "created_at": "2026-09-21T00:02:25Z", + "updated_at": "2026-09-21T00:02:26Z", + "homepage": null, + "default_branch": "main", + "default_sha": null, + "readme": null, + "topics": [], + "desc_hash": null + }, + { + "id": "Xubqpanda/JevLoop", + "description": "The agent loop where decisions don't cost a model call. Zero deps, runs offline, no API key needed.", + "stars": 0, + "forks": 0, + "license": "Apache-2.0", + "language": "TypeScript", + "size": 150, + "pushed_at": "2026-09-21T01:46:46Z", + "created_at": "2026-09-20T15:21:09Z", + "updated_at": "2026-09-21T01:46:50Z", + "homepage": "https://github.com/Xubqpanda/JevRepo", + "default_branch": "main", + "default_sha": "50236bf2221bdc9b3f556ad4f6e067daf558c12b", + "readme": { + "path": "README.md", + "sha": "0e3f67c72801630bee7149c7e8157ea1370c37fa", + "size": 11730 + }, + "topics": [ + "agent-framework", + "agent-loop", + "ai-agent", + "decision-model", + "jev", + "llm", + "system-one", + "typescript", + "zero-dependency" + ], + "desc_hash": "ab62fbd74fcf" + }, + { + "id": "YuanKJing/Jev-as-Policy", + "description": "The highly anticipated open-source repository for JEV as Policy enables one-click setup of the simulation environment. Evaluations of Astra + JEV on benchmarks such as RoboTwin will also be released soon.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:43:07Z", + "created_at": "2026-09-21T01:43:06Z", + "updated_at": "2026-09-21T01:48:09Z", + "homepage": "", + "default_branch": "main", + "default_sha": "c2e1e17b6bb05fdbea1284d00730597cca161443", + "readme": { + "path": "README.md", + "sha": "2b83b5db78bd8d1bf4acdeb90fe8ab8e04b7948f", + "size": 225 + }, + "topics": [], + "desc_hash": "86f0b2c6ad77" + }, + { + "id": "Zafer-Liu/jev-demos", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": "HTML", + "size": 0, + "pushed_at": "2026-09-21T01:14:29Z", + "created_at": "2026-09-21T01:14:24Z", + "updated_at": "2026-09-21T01:14:33Z", + "homepage": null, + "default_branch": "main", + "default_sha": "c875343d3780a79426854e7795abf0e63f7c3dfa", + "readme": { + "path": "README.md", + "sha": "b9976762011fbb55fa71f734f2a0da2c2a4fae5e", + "size": 3656 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "aboisvert/jevvy", + "description": "Use jev model to augment csv files with inferred classification, scoring, or probability scores", + "stars": 0, + "forks": 0, + "license": "Apache-2.0", + "language": "Scala", + "size": 38, + "pushed_at": "2026-09-21T01:26:21Z", + "created_at": "2026-09-21T00:51:25Z", + "updated_at": "2026-09-21T01:23:09Z", + "homepage": "https://github.com/aboisvert/jevvy", + "default_branch": "main", + "default_sha": "14de871f75034a217d2d9aa2567e3ff5e1445a5c", + "readme": { + "path": "README.md", + "sha": "939119dd6c04d66d820f85a1ee2b59cb79962339", + "size": 4592 + }, + "topics": [ + "classification", + "csv-processing", + "graalvm-native-image", + "jev", + "probability", + "scala", + "scoring" + ], + "desc_hash": "77d47c2ad02b" + }, + { + "id": "andrest04/jev-lab", + "description": "Local lab for learning and testing TypeSafe Jev (System One)", + "stars": 0, + "forks": 0, + "license": null, + "language": "JavaScript", + "size": 0, + "pushed_at": "2026-09-21T01:24:20Z", + "created_at": "2026-09-21T01:23:03Z", + "updated_at": "2026-09-21T01:24:21Z", + "homepage": null, + "default_branch": "main", + "default_sha": "3ef5eb16c288bb41009d1c3174e7b0a6e1225b3e", + "readme": { + "path": "README.md", + "sha": "741995e66b02d6a11aa07c40620189ced9d3a3e8", + "size": 2063 + }, + "topics": [], + "desc_hash": "a838c0156f8c" + }, + { + "id": "andyrewlee/awesome-system-one", + "description": "Curated list of tools related to system one models", + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 9, + "pushed_at": "2026-09-20T07:36:47Z", + "created_at": "2026-09-19T23:20:48Z", + "updated_at": "2026-09-20T07:36:51Z", + "homepage": null, + "default_branch": "main", + "default_sha": "1345d8d237d33b5f3a43e99899741a2cb090df41", + "readme": { + "path": "README.md", + "sha": "5bf4d8b34e084d9861a80c4196abe4533276ac2d", + "size": 11919 + }, + "topics": [], + "desc_hash": "77976d967f10" + }, + { + "id": "ashafizullah/jev-triage", + "description": "Automated issue & PR triage for open-source maintainers, powered by Jev (TypeSafe AI).", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "TypeScript", + "size": 0, + "pushed_at": "2026-09-21T01:45:25Z", + "created_at": "2026-09-21T00:09:58Z", + "updated_at": "2026-09-21T01:45:28Z", + "homepage": null, + "default_branch": "main", + "default_sha": "6907ccc444a25c02f58f6f3dfaeec281bec76fa6", + "readme": { + "path": "README.md", + "sha": "96dcb0a587985b7c90bac63b25a28b625c1b3643", + "size": 16627 + }, + "topics": [], + "desc_hash": "9dead8b5a1bc" + }, + { + "id": "blanket11/jev-guide-ja", + "description": "jev\u306e\u5b66\u7fd2", + "stars": 0, + "forks": 0, + "license": null, + "language": "JavaScript", + "size": 0, + "pushed_at": "2026-09-21T01:37:48Z", + "created_at": "2026-09-21T01:15:10Z", + "updated_at": "2026-09-21T01:37:51Z", + "homepage": null, + "default_branch": "main", + "default_sha": "d844446d83867f0a93b72c74a1ded11ad130ee6a", + "readme": { + "path": "README.md", + "sha": "b6702287356c89aa2fe0e2735e8ac062897a1a19", + "size": 8518 + }, + "topics": [], + "desc_hash": "7006ad0d6b76" + }, + { + "id": "caohy1988/jev-guard-smoke", + "description": "Lab smoke proof for leepokai/jev-guard (Awesome-Jev #1). No API keys.", + "stars": 0, + "forks": 0, + "license": null, + "language": "Shell", + "size": 0, + "pushed_at": "2026-09-21T01:47:42Z", + "created_at": "2026-09-21T01:42:42Z", + "updated_at": "2026-09-21T01:42:52Z", + "homepage": null, + "default_branch": "docs/jev-guard-smoke-lab", + "default_sha": "8d3f629e1cb20ef55e2c379cca3379625f45d609", + "readme": null, + "topics": [], + "desc_hash": "3ad3b889d119" + }, + { + "id": "dinkarjuyal/jev-gepa", + "description": "Wiring a fast local NLI judge (Jev) into GEPA's reflective prompt optimization loop", + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:34:10Z", + "created_at": "2026-09-21T01:34:06Z", + "updated_at": "2026-09-21T01:34:14Z", + "homepage": null, + "default_branch": "main", + "default_sha": "3d52ba9b22397cb827b4f682db465ddf7d9aaffb", + "readme": { + "path": "README.md", + "sha": "655920a0144af6584d044638a77b83f3c81c1cde", + "size": 6297 + }, + "topics": [], + "desc_hash": "c4e769e2e075" + }, + { + "id": "early-effect/hexis", + "description": "ZIO / Scala 3 SDK for TypeSafe System One (Jev)", + "stars": 0, + "forks": 0, + "license": "Apache-2.0", + "language": "Scala", + "size": 0, + "pushed_at": "2026-09-21T01:54:48Z", + "created_at": "2026-09-21T01:50:01Z", + "updated_at": "2026-09-21T01:54:53Z", + "homepage": "https://www.earlyeffect.rocks/hexis/", + "default_branch": "main", + "default_sha": "c2d622c130baed8876f8f7fdca23e8d33c83deec", + "readme": { + "path": "README.md", + "sha": "4ffae17a7060ebc510508b4b5aab48b475fdcc96", + "size": 1468 + }, + "topics": [], + "desc_hash": "9c215f1aee74" + }, + { + "id": "elberacasa/omawish", + "description": "A System One for Omarchy: type what you want, and your desktop does it. Local, instant, 33M parameters, fine-tuned on one gaming GPU.", + "stars": 0, + "forks": 0, + "license": "NOASSERTION", + "language": "Rust", + "size": 0, + "pushed_at": "2026-09-21T01:48:04Z", + "created_at": "2026-09-21T01:34:23Z", + "updated_at": "2026-09-21T01:51:59Z", + "homepage": "https://github.com/elberacasa/omawish#readme", + "default_branch": "main", + "default_sha": "82c43e5425df4acf75459a9350115d0a849c5f0b", + "readme": { + "path": "README.md", + "sha": "bb2be38501638237b1b16ea70f1302b48b78861d", + "size": 17950 + }, + "topics": [ + "command-palette", + "hyprland", + "linux-desktop", + "local-first", + "omarchy", + "quickshell", + "rust", + "sentence-transformers", + "system-one" + ], + "desc_hash": "e53948cca959" + }, + { + "id": "endman100/research-Qwen3.8-JevLike", + "description": "71-label binary routing vs JSON Schema on Qwen3.8 NVFP4 / RTX 5090: measurements, raw evidence and ideal-parallel analysis", + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:31:38Z", + "created_at": "2026-09-21T01:18:14Z", + "updated_at": "2026-09-21T01:31:42Z", + "homepage": null, + "default_branch": "main", + "default_sha": "b406b17c3936e23ba588943cf05a15b8b27e0943", + "readme": { + "path": "README.md", + "sha": "7c908289d0e8a5ba36dbef64034ecf281dd2d89a", + "size": 5354 + }, + "topics": [], + "desc_hash": "90a2bc4aa36a" + }, + { + "id": "eteen12/jev-browser-automation", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:14:02Z", + "created_at": "2026-09-21T01:14:01Z", + "updated_at": "2026-09-21T01:14:02Z", + "homepage": null, + "default_branch": "main", + "default_sha": null, + "readme": null, + "topics": [], + "desc_hash": null + }, + { + "id": "fruitymcdoo/JevChat", + "description": "A chat interface built on Jev, TypeSafe's decision-only model: every word is a typed decision", + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:56:41Z", + "created_at": "2026-09-21T01:40:58Z", + "updated_at": "2026-09-21T01:56:45Z", + "homepage": null, + "default_branch": "main", + "default_sha": "6e2bc022cbbec86f055b57f6e67195642390434a", + "readme": { + "path": "README.md", + "sha": "abb03d83190aedc8055dda09aa1f79aa7dd2ef51", + "size": 11345 + }, + "topics": [], + "desc_hash": "dcde87688f6c" + }, + { + "id": "ismaelsoilet/jev-harness", + "description": null, + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:49:05Z", + "created_at": "2026-09-21T01:48:13Z", + "updated_at": "2026-09-21T01:49:08Z", + "homepage": null, + "default_branch": "main", + "default_sha": "9686bfd363c7f3bf818ea94c474e4df5e9e4e939", + "readme": { + "path": "README.md", + "sha": "7d6a0078ec5f306dec87b1508d1292dbdeda6f95", + "size": 12696 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "jacks3tr/Jev-Desktop", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:27:39Z", + "created_at": "2026-09-21T01:27:39Z", + "updated_at": "2026-09-21T01:27:39Z", + "homepage": null, + "default_branch": "main", + "default_sha": null, + "readme": null, + "topics": [], + "desc_hash": null + }, + { + "id": "javsanesq/jevlab", + "description": "A terminal workbench for learning, testing, and connecting TypeSafe Jev decisions", + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 548, + "pushed_at": "2026-09-21T00:58:07Z", + "created_at": "2026-09-21T00:09:01Z", + "updated_at": "2026-09-21T00:53:43Z", + "homepage": null, + "default_branch": "main", + "default_sha": "df7198de170086c2714132fdbd67061b4cfb764f", + "readme": { + "path": "README.md", + "sha": "87d810d6a4b2633840606f0c18dd2313feaaeff2", + "size": 30431 + }, + "topics": [], + "desc_hash": "65f3b669a7ca" + }, + { + "id": "joelakaufmann-lgtm/NRS-Navigator", + "description": "Local Nevada statute search and a reproducible evaluation of Jev-assisted ranking against keyword search.", + "stars": 0, + "forks": 0, + "license": "Apache-2.0", + "language": "HTML", + "size": 0, + "pushed_at": "2026-09-21T01:45:20Z", + "created_at": "2026-09-21T01:45:05Z", + "updated_at": "2026-09-21T01:45:26Z", + "homepage": null, + "default_branch": "main", + "default_sha": "6c9394e6c6cb78b4c0df6371a8b48eea5f6b47b9", + "readme": { + "path": "README.md", + "sha": "dbe9da1004bdce5116f5aa9cee46ead9b40e37bd", + "size": 17445 + }, + "topics": [], + "desc_hash": "5c3f19d9c202" + }, + { + "id": "k-srkw/jev-playground", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:38:24Z", + "created_at": "2026-09-21T01:25:57Z", + "updated_at": "2026-09-21T01:38:22Z", + "homepage": null, + "default_branch": "main", + "default_sha": "d935d01d70ab8c23a500f8006d57928124def021", + "readme": { + "path": "README.md", + "sha": "0b0945882b0c32c891aa354d9209ac5561ea9c19", + "size": 16 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "koteitan/laya-bot-det", + "description": "bot detector by laya for nostr", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": null, + "size": 1, + "pushed_at": "2026-09-21T00:57:03Z", + "created_at": "2026-09-21T00:57:02Z", + "updated_at": "2026-09-21T00:57:07Z", + "homepage": null, + "default_branch": "main", + "default_sha": "93120b7f808abe98c623ea4a7c9ebae16c809aa4", + "readme": null, + "topics": [], + "desc_hash": "c3f47e577eed" + }, + { + "id": "kzkhykw/jev-or-not", + "description": "Jev\u308b\uff1f\u30e2\u30c7\u30eb\u9078\u629e\u30d5\u30ed\u30fc\u30c1\u30e3\u30fc\u30c8", + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:13:53Z", + "created_at": "2026-09-21T01:04:37Z", + "updated_at": "2026-09-21T01:13:57Z", + "homepage": null, + "default_branch": "jev", + "default_sha": "916a83d68c9ae8ef4232c3ae86a1aee45c695a06", + "readme": { + "path": "README.md", + "sha": "7cbe3a508fab5775105262379189e6f0011722d2", + "size": 1120 + }, + "topics": [], + "desc_hash": "a3279deb950a" + }, + { + "id": "laidick/system-one-benchmark", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 14, + "pushed_at": "2026-09-19T23:48:40Z", + "created_at": "2026-09-19T23:48:35Z", + "updated_at": "2026-09-19T23:48:44Z", + "homepage": null, + "default_branch": "main", + "default_sha": "ccd4b79d1a516420450cd0811b86fd7aa528ef05", + "readme": { + "path": "README.md", + "sha": "5ffe72671473b1a0462e0dc7f05710e040eedc46", + "size": 10650 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "lezgoverci/jev-docs", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 410, + "pushed_at": "2026-09-21T00:04:09Z", + "created_at": "2026-09-21T00:04:01Z", + "updated_at": "2026-09-21T00:04:13Z", + "homepage": null, + "default_branch": "main", + "default_sha": "db9c9a85d63d2f0bae1d52856716a6a7f1b1d2e9", + "readme": { + "path": "README.md", + "sha": "f06bc61d36e9db7d4677fde855cc2bdd710eadf1", + "size": 14802 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "luckberonne/mini-jev", + "description": "Clasificador de comandos de shell de una sola pasada (solo lectura / reversible / destructivo), inspirado en Jev", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 54, + "pushed_at": "2026-09-21T00:01:26Z", + "created_at": "2026-09-21T00:01:21Z", + "updated_at": "2026-09-21T00:01:30Z", + "homepage": null, + "default_branch": "master", + "default_sha": "97d4b30b6f5de282c66809c4a34c5ea164454995", + "readme": { + "path": "README.md", + "sha": "795ab56ea245d37dde4c362999a6cc9a42c7d602", + "size": 2630 + }, + "topics": [], + "desc_hash": "06269d8dace1" + }, + { + "id": "marcosmartinez/jev-acento", + "description": "\u00bfJev entiende tu acento? Pre-registered audit of TypeSafe AI's Jev on Spanish \u2014 accuracy, calibration and token cost \u2014 plus a CLI to run the same comparison on your own labelled data.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 2285, + "pushed_at": "2026-09-21T01:12:19Z", + "created_at": "2026-09-20T23:23:21Z", + "updated_at": "2026-09-21T01:12:46Z", + "homepage": "https://github.com/marcosmartinez/jev-acento/releases/latest", + "default_branch": "main", + "default_sha": "7e007b4c2bd552471b56eff035f9a3df7593f9fe", + "readme": { + "path": "README.md", + "sha": "994943244b6ba1fdaf3420b2dcf5a9066f158509", + "size": 12639 + }, + "topics": [ + "benchmark", + "calibration", + "expected-calibration-error", + "jev", + "llm-evaluation", + "nlp", + "pre-registration", + "reproducible-research", + "spanish-nlp", + "typesafe-ai" + ], + "desc_hash": "1f8cc7248824" + }, + { + "id": "mednabouli/jev-ai-polymarket-copy-trading", + "description": "Automated Polymarket copy trading bot with MCP servers, Telegram alerts, and profitable wallet tracking. Zero API keys - uses Claude Code OAuth session auth.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 174, + "pushed_at": "2026-09-21T01:59:45Z", + "created_at": "2026-09-20T23:51:24Z", + "updated_at": "2026-09-21T01:59:48Z", + "homepage": null, + "default_branch": "main", + "default_sha": "aad77195fff5e1b4dafc9474a2f21239c824f33e", + "readme": { + "path": "README.md", + "sha": "7bb1a24b410018502e3263e43b1646b24008426e", + "size": 4581 + }, + "topics": [], + "desc_hash": "7082b3ecee24" + }, + { + "id": "olivdx/jev-mcp", + "description": "Jev-powered decision layer for coding agents. Analyze code and diffs, assess bugs, security, risk, and breaking changes, and return structured decisions for automated continue, fix, retry, or human-review workflows.", + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:42:36Z", + "created_at": "2026-09-21T01:42:35Z", + "updated_at": "2026-09-21T01:42:36Z", + "homepage": null, + "default_branch": "main", + "default_sha": null, + "readme": null, + "topics": [], + "desc_hash": "418b7804fdf5" + }, + { + "id": "pattoor/JEV-agent-opencv", + "description": "Juego simple para probar el modelo JEV con vision", + "stars": 0, + "forks": 0, + "license": null, + "language": null, + "size": 0, + "pushed_at": "2026-09-21T01:18:22Z", + "created_at": "2026-09-21T01:18:21Z", + "updated_at": "2026-09-21T01:18:22Z", + "homepage": null, + "default_branch": "main", + "default_sha": null, + "readme": null, + "topics": [], + "desc_hash": "9b5f38d0b031" + }, + { + "id": "peach-zhang/typesafe-go", + "description": "TypeSafe System One (Jev) \u7684 Go SDK \u2014 \u7c7b\u578b\u5316\u5224\u65ad\u4e0e\u6982\u7387,\u4ee3\u7801\u638c\u63a7\u5de5\u4f5c\u6d41", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Go", + "size": 0, + "pushed_at": "2026-09-21T01:42:10Z", + "created_at": "2026-09-21T01:34:16Z", + "updated_at": "2026-09-21T01:42:13Z", + "homepage": null, + "default_branch": "main", + "default_sha": "cd4b6bf028f73244440b898c5704b33bfa95f7af", + "readme": { + "path": "README.md", + "sha": "109ec34a82475c67df79d5ad2400d1314662d8e9", + "size": 6520 + }, + "topics": [], + "desc_hash": "9dd9a4e347ce" + }, + { + "id": "promptgtm-shared/clay-jev-people-ranker", + "description": "Agent Skill and Python workflow for Clay lead scoring, B2B prospect qualification, and people-search ranking with TypeSafe JEV.", + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:22:20Z", + "created_at": "2026-09-21T01:15:47Z", + "updated_at": "2026-09-21T01:22:23Z", + "homepage": null, + "default_branch": "main", + "default_sha": "2494b65218d4e4e940911db7c485710a3cff7050", + "readme": { + "path": "README.md", + "sha": "40402cc861cb84d1ace9110cabe24405505dd536", + "size": 12210 + }, + "topics": [ + "agent-skills", + "ai-agents", + "ai-lead-scoring", + "b2b-sales", + "claude-code", + "clay", + "clay-cli", + "cursor", + "grok", + "gtm-engineering", + "jev", + "lead-qualification", + "lead-scoring", + "people-search", + "prospecting", + "python", + "revops", + "sales-automation", + "typesafe-ai", + "typesafe-jev" + ], + "desc_hash": "1d3275407722" + }, + { + "id": "sahasrarjn/system-one", + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 62, + "pushed_at": "2026-09-20T01:51:11Z", + "created_at": "2026-09-19T06:07:11Z", + "updated_at": "2026-09-20T01:51:15Z", + "homepage": null, + "default_branch": "main", + "default_sha": "bb2383834d198ca00ffb498c063ba1b455e1ff87", + "readme": { + "path": "README.md", + "sha": "b71d9004231903688f450731977d5a482ff777ae", + "size": 8044 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "sarathi-aiml/jevsql", + "description": "Text-to-SQL where the model never writes SQL \u2014 typed, calibrated decisions (TypeSafe Jev) + code-assembled queries", + "stars": 0, + "forks": 0, + "license": null, + "language": "Python", + "size": 0, + "pushed_at": "2026-09-21T01:50:41Z", + "created_at": "2026-09-21T01:28:29Z", + "updated_at": "2026-09-21T01:50:45Z", + "homepage": null, + "default_branch": "main", + "default_sha": "ba46b8b400a1e13dfa468bbfce2f370cf9d01972", + "readme": { + "path": "README.md", + "sha": "9c05b06bf11aac2ed94e953eb71fcfd415e56c5c", + "size": 10324 + }, + "topics": [], + "desc_hash": "b5cb26f386df" + }, + { + "id": "theosunny/jev_stock", + "description": null, + "stars": 0, + "forks": 0, + "license": "MIT", + "language": "Python", + "size": 54, + "pushed_at": "2026-09-21T01:48:42Z", + "created_at": "2026-09-20T09:11:00Z", + "updated_at": "2026-09-21T01:48:45Z", + "homepage": null, + "default_branch": "main", + "default_sha": "0c0e3fbbb7861432e70a724036e4b26e12f03e42", + "readme": { + "path": "README.md", + "sha": "063036893cb5832a65843187a7373b734535af9a", + "size": 3924 + }, + "topics": [], + "desc_hash": null + }, + { + "id": "twilwa/pi-typesafe", + "description": "Pi coding-agent extension built on the TypeSafe AI System One API (Jev)", + "stars": 0, + "forks": 0, + "license": null, + "language": "TypeScript", + "size": 94, + "pushed_at": "2026-09-21T01:43:14Z", + "created_at": "2026-09-17T00:18:58Z", + "updated_at": "2026-09-21T01:43:18Z", + "homepage": null, + "default_branch": "main", + "default_sha": "3dc40b969dc3c8f90e313622c3e69a91b63b7cc6", + "readme": { + "path": "README.md", + "sha": "195563911db2f04799efe12a707a66e695eb46c3", + "size": 20418 + }, + "topics": [], + "desc_hash": "d86a28e2715b" + }, + { + "id": "willgriffin/pi-fusion-matrix", + "description": "Multi-model deliberation for the pi coding agent: named fusions, per-slot fallback and routing, version-free aliases, conservative decision backends", + "stars": 0, + "forks": 0, + "license": null, + "language": "JavaScript", + "size": 592, + "pushed_at": "2026-09-21T01:46:54Z", + "created_at": "2026-09-18T18:53:12Z", + "updated_at": "2026-09-21T01:46:58Z", + "homepage": null, + "default_branch": "main", + "default_sha": "f5d2a1b89448bd2dd84ae563a8f1d5b613e11425", + "readme": { + "path": "README.md", + "sha": "dfb8b07ad2c06d212db224e664601ec9bad42958", + "size": 49676 + }, + "topics": [], + "desc_hash": "0fadc538568e" + } +] diff --git a/research/archive/hourly/2026-09-21T01/prompt_Augustus.md b/research/archive/hourly/2026-09-21T01/prompt_Augustus.md new file mode 100644 index 0000000..211dd87 --- /dev/null +++ b/research/archive/hourly/2026-09-21T01/prompt_Augustus.md @@ -0,0 +1,102 @@ +# Fold hourly 1946 HIGH → Augustus + +Boise label **1946** (~2026-09-20 19:46 MDT / fired 2026-09-21T01:46:45Z). +Mode: **NEW_OFF_MAIN** (prior 1843 fold PRs merged; do not reply finished agents). +Repo: `24601/Augustus` + +## Items (see list; Open-Jev deferred to #53 densify) + +- `brainstormity/Jev-X-Sentiment-Analysis` [novel] stars=135 — +- `ikermoel/open-alternative-jev` [novel] stars=36 — Open-source alternative to TypeSafe's Jev: a System One style model layer that gives typed, calibrated decisions from any open-weights LLM in one forward pass (HF + vLLM), with hon +- `okinaaudio/live-jev` [novel] stars=36 — Control Ableton Live with one short sentence (Japanese / English). Summon with ⌘⇧Space, type or dictate, done. +- `heyjunpenn/awesome-jev` [novel] stars=32 — A verified, community-maintained catalog of 503 open-source projects built with Jev. +- `NanmiCoder/jev-arena` [novel] stars=31 — Jev 模型介绍与实测：通过 Choice / Score / Noul 将自然语言转为带类型的判断与概率，用于分类、评分和路由；支持与 DeepSeek 等模型对比评论打标、速度与结果，含 CSV/Excel 导入、原速回放与离线报告。 +- `yzfly/awesome-jev-zh` [novel] stars=31 — Jev / TypeSafe System One 中文精选列表：官方资料、SDK、爆款应用、Agent 工具、开源复现与独立评测，附中文上手指南，每日自动收录 GitHub 热门项目。 +- `openroboto-ai/jev-robot-control` [novel] stars=27 — +- `win4r/jev-skill-suggester` [novel] stars=27 — 用 TypeSafe Jev 推荐已安装 Skill / Bounded installed-skill recommendations with TypeSafe Jev. Python CLI, Codex skill, bilingual docs and live examples. +- `PyModel/typesafe-mcp` [novel] stars=23 — +- `AkashPriyadarshii/jev-seo` [novel] stars=21 — 100% free ₹0 agent-first SEO & GEO CLI suite and MCP server in Rust replacing Semrush and OpenSEO via DuckDuckGo and TypeSafe Jev System One +- `devtooligan/jevscan-evm` [novel] stars=21 — +- `mizzlelover/jev-hub` [novel] stars=21 — JEV HUB · X 上关于 TypeSafe AI「系统一模型」Jev 的长文与演示视频聚合（保留原链与作者）｜ 谁是专家 出品 +- `sgoedecke/system-one` [novel] stars=20 — Batched single-token choice inference for open language models, compatible with TypeSafe +- `zszz3/Pi-Jev-Guide` [novel] stars=19 — +- `mithalouni/system-one-open` [novel] stars=18 — Open replica of TypeSafe's Jev: typed calibrated decisions in one forward pass, on Gemma 4 E2B / Gemma 3 270M (Modal) +- `davila7/jev-explained` [novel] stars=14 — Jev Explained +- `ckaraca/awesome-jev` [novel] stars=7 — A curated list of tools, integrations, and experiments built on Jev, TypeSafe AI's System One model for fast, typed decisions. +- `arunav25/jev-mcp` [novel] stars=5 — Connect JEV to MCP clients and compare its judgments against general-purpose LLMs using shared datasets and measurable accuracy. +- `cobusgreyling/Jev` [novel] stars=3 — Unofficial TypeSafe Jev showcase — System One decisions, not chat. +- `nrdz-labs/fast-jev-opencode` [novel] stars=3 — Jev-scored context pruning for OpenCode: drops stale tool calls and truncates bulky results on the outgoing request — fail-open, cache-backed, configurable live. Port of fast-jev-c +- `liao96312/jev-arena-nanojev` [novel] stars=2 — 完全本地的 NanoJev 网格决策游戏实验场，支持中文 Pygame、多关卡与 GTX 1660S 训练 +- `Krug2/JevLM-Open` [novel] stars=1 — +- `Rizzo-AI-Academy/rizzo-flow` [novel] stars=1 — The open, local take on Jev: typed decisions from an LLM, without generating a single token +- `clouatre-labs/decisions-judge-mcp` [novel] stars=1 — MCP server exposing an LLM judge (typed decisions: yes/no probability, choice, score) for coding agents +- `emirbartu/opencode-system-one` [novel] stars=1 — Opencode plugin using Jev (system one model) as part of software development process. Not affiliated with Opencode team. +- `kotoba-lang/typed-decisions` [novel] stars=1 — Jev-shaped typed-decision model (state + Choice/Score/Noul questions -> calibrated probabilities, one pass) on ModernBERT / DeBERTa / LLaDA-MoE, with measured latency, accuracy, ca +- `mallahyari/system-one-benchmark` [novel] stars=1 — +- `sable-inc/jev-linter-action` [novel] stars=1 — Configurable semantic CI checks for repository files using TypeSafe Jev +- `shirenchuang/awsomejev` [novel] stars=1 — Awesome Jev：Jev 开源生态导航 +- `1816586742-stack/jev-craft` [novel] stars=0 — 让 Agent 长出「操作杆」：用 System One 模型（Jev）承担高频判断、多模态模型当眼睛。给思路 + 可跑的参考实现（75 条离线断言，零依赖零 key） +- `DolphinMiner/jev-rss` [novel] stars=0 — A local-first RSS reader with Jev-powered semantic screening. Follow what matters, inspect every judgment, and keep control of your reading. English / 简体中文. +- `Dreydrey9000/jev-relay` [novel] stars=0 — Guarded local-first decision advice for Claude Code, Codex and Hermes/Jax, with explicit Jev checks and review fallbacks. +- `Eric-Zhou-0302/jev-A-share-trader` [novel] stars=0 — A Jev-powered technical analysis workspace for China A-shares, supporting AKShare/Tushare, market scans, and Buy/Hold/Sell assessments with time horizons and traceable evidence. +- `JTech-CO/Jev-Simulink-Supervisor` [novel] stars=0 — Jev-Simulink Supervisor +- `Kwwwww74/OpenJev` [novel] stars=0 — +- `RuipuCui/jev-harness` [novel] stars=0 — +- `Xubqpanda/JevLoop` [novel] stars=0 — The agent loop where decisions don't cost a model call. Zero deps, runs offline, no API key needed. +- `YuanKJing/Jev-as-Policy` [novel] stars=0 — The highly anticipated open-source repository for JEV as Policy enables one-click setup of the simulation environment. Evaluations of Astra + JEV on benchmarks such as RoboTwin wil +- `Zafer-Liu/jev-demos` [novel] stars=0 — +- `aboisvert/jevvy` [novel] stars=0 — Use jev model to augment csv files with inferred classification, scoring, or probability scores +- `andrest04/jev-lab` [novel] stars=0 — Local lab for learning and testing TypeSafe Jev (System One) +- `andyrewlee/awesome-system-one` [novel] stars=0 — Curated list of tools related to system one models +- `ashafizullah/jev-triage` [novel] stars=0 — Automated issue & PR triage for open-source maintainers, powered by Jev (TypeSafe AI). +- `blanket11/jev-guide-ja` [novel] stars=0 — jevの学習 +- `caohy1988/jev-guard-smoke` [novel] stars=0 — Lab smoke proof for leepokai/jev-guard (Awesome-Jev #1). No API keys. +- `dinkarjuyal/jev-gepa` [novel] stars=0 — Wiring a fast local NLI judge (Jev) into GEPA's reflective prompt optimization loop +- `early-effect/hexis` [novel] stars=0 — ZIO / Scala 3 SDK for TypeSafe System One (Jev) +- `elberacasa/omawish` [novel] stars=0 — A System One for Omarchy: type what you want, and your desktop does it. Local, instant, 33M parameters, fine-tuned on one gaming GPU. +- `endman100/research-Qwen3.8-JevLike` [novel] stars=0 — 71-label binary routing vs JSON Schema on Qwen3.8 NVFP4 / RTX 5090: measurements, raw evidence and ideal-parallel analysis +- `eteen12/jev-browser-automation` [novel] stars=0 — +- `fruitymcdoo/JevChat` [novel] stars=0 — A chat interface built on Jev, TypeSafe's decision-only model: every word is a typed decision +- `ismaelsoilet/jev-harness` [novel] stars=0 — +- `jacks3tr/Jev-Desktop` [novel] stars=0 — +- `javsanesq/jevlab` [novel] stars=0 — A terminal workbench for learning, testing, and connecting TypeSafe Jev decisions +- `joelakaufmann-lgtm/NRS-Navigator` [novel] stars=0 — Local Nevada statute search and a reproducible evaluation of Jev-assisted ranking against keyword search. +- `k-srkw/jev-playground` [novel] stars=0 — +- `koteitan/laya-bot-det` [novel] stars=0 — bot detector by laya for nostr +- `kzkhykw/jev-or-not` [novel] stars=0 — Jevる？モデル選択フローチャート +- `laidick/system-one-benchmark` [novel] stars=0 — +- `lezgoverci/jev-docs` [novel] stars=0 — +- `luckberonne/mini-jev` [novel] stars=0 — Clasificador de comandos de shell de una sola pasada (solo lectura / reversible / destructivo), inspirado en Jev +- `marcosmartinez/jev-acento` [novel] stars=0 — ¿Jev entiende tu acento? Pre-registered audit of TypeSafe AI's Jev on Spanish — accuracy, calibration and token cost — plus a CLI to run the same comparison on your own labelled da +- `mednabouli/jev-ai-polymarket-copy-trading` [novel] stars=0 — Automated Polymarket copy trading bot with MCP servers, Telegram alerts, and profitable wallet tracking. Zero API keys - uses Claude Code OAuth session auth. +- `olivdx/jev-mcp` [novel] stars=0 — Jev-powered decision layer for coding agents. Analyze code and diffs, assess bugs, security, risk, and breaking changes, and return structured decisions for automated continue, fix +- `pattoor/JEV-agent-opencv` [novel] stars=0 — Juego simple para probar el modelo JEV con vision +- `peach-zhang/typesafe-go` [novel] stars=0 — TypeSafe System One (Jev) 的 Go SDK — 类型化判断与概率,代码掌控工作流 +- `promptgtm-shared/clay-jev-people-ranker` [novel] stars=0 — Agent Skill and Python workflow for Clay lead scoring, B2B prospect qualification, and people-search ranking with TypeSafe JEV. +- `sahasrarjn/system-one` [novel] stars=0 — +- `sarathi-aiml/jevsql` [novel] stars=0 — Text-to-SQL where the model never writes SQL — typed, calibrated decisions (TypeSafe Jev) + code-assembled queries +- `theosunny/jev_stock` [novel] stars=0 — +- `twilwa/pi-typesafe` [novel] stars=0 — Pi coding-agent extension built on the TypeSafe AI System One API (Jev) +- `willgriffin/pi-fusion-matrix` [novel] stars=0 — Multi-model deliberation for the pi coding agent: named fusions, per-slot fallback and routing, version-free aliases, conservative decision backends +- `alexwestco/llm-to-jev` [revisit] stars=3 — REVISIT: description rewrite 'llm-to-jev' → 'Convert LLM prompts to Jev prompts' (desc_hash change); SHA unchanged 234058ab372d + +## Instructions +- Fold into catalog/class table / recipes as appropriate for this repo's role. +- Mark third-party benches *theirs*. +- Human-facing prose: avoid AI tells (em dashes etc.). +- Augustus: design judgment, measurement, class table, decision-model benefit recipes. +- Jev-omni: omni/ports/kits/community. +- rh-guard: gate/router/hook cousins only. + +## Coordination note (parent) +- Skip densify for `Zefan-Cai/Open-Jev`: Augustus #53 and a Jev-omni densify agent are already running on that densify. Do not open a competing Open-Jev densify card; focus on novel HIGH + `alexwestco/llm-to-jev` description densify. +- Augustus #54 (aisearchio census gaps) may already cover `sgoedecke/system-one`, `mithalouni/system-one-open`, `kotoba-lang/typed-decisions`. If cards exist on main or in #54, densify lightly or skip duplicates rather than forking parallel cards. + + +## Success criteria +1. Open ONE PR off main (NEW_OFF_MAIN). Title like: Fold hourly 1946 HIGH +2. Add/update class-table / catalog entries for novel HIGH items; light densify for alexwestco/llm-to-jev description rewrite. +3. Decision-model benefit recipes where Augustus role applies (not Jev-only). +4. Mark third-party benches *theirs*. Human-facing prose: no em dashes / AI tells. +5. Do NOT densify Zefan-Cai/Open-Jev (covered by open #53). +6. Report PR URL when done. diff --git a/research/archive/hourly/2026-09-21T01/readmes.json b/research/archive/hourly/2026-09-21T01/readmes.json new file mode 100644 index 0000000..6c2e5d9 --- /dev/null +++ b/research/archive/hourly/2026-09-21T01/readmes.json @@ -0,0 +1,281 @@ +{ + "brainstormity/Jev-X-Sentiment-Analysis": { + "sha": "bf4134b44cda73599f005e9408391353a2ed9437", + "path": "README.md", + "size": 7370, + "excerpt": "# Jev X Sentiment Analysis\n![Jev X Sentiment Analysis](public/images/frontend.jpeg)\nAn on-demand crypto market intelligence and decision-support terminal powered by TypeSafe AI's System One model (Jev).\nThe tool allows you to search any cryptocurrency (such as BTC, SOL, or ETH), select how many tweets you want to analyze (from 50 up to 1,000 tweets), and receive an instant, data-backed trading decision (Buy, Sell, Hold, or Take Profit) based on real-time market data, perpetuals funding rates, and social sentiment.\nThe platform does not execute trades automatically. It generates a clear decision card with calculated entry ranges, stop losses, and target levels so you can review the reasoning and execute manually on whichever exchange or DEX you prefer.\n---\n## How Tweets Are Analyzed\nWhen you request sentiment analysis across 100 to 1,000 tweets, dumping hundreds of raw tweets into an LLM would exceed token budgets and introduce latency. Instead, Jev X Sentiment Analysis uses an **intelligent two-tier pipeline**:\n```text\nUser Search (e.g. \"SOL\", 500 tweets)\n │\n ├──> CCXT: Live Price, 24h Volume, RSI, Funding Rate, Open Interest\n └──> TwitterAPI.io: 500 tweets ingested via cursor pagination\n │\n ▼\n Tier 1: Python Statistical Pre-Processing\n - Computes total engagement velocity (likes, retweets per minute)\n - Measures author diversity ratio (detects bot farms vs organic retail)\n - Computes keyword sentiment polarity (fear/capitulation vs greed/hype)\n - Stratified extraction:\n • Top 25 highest-engaged tweets (KOL & market-moving opinions)\n • 25 most recent breaking tweets (current real-time narrative)\n │\n ▼\n Tier 2: Early-Stopping Database Deduplication (API Cost Optimization)\n - To save API credits, the system stores all ingested tweets in a local SQLite database (`data/market_intel.db`).\n - TwitterAPI.io returns tweets in reverse-chronological order (`queryType=\"Latest\"`).\n - During pagination, **as soon as a returned tweet ID already exists in the local database, the pagination loop halts immediately**.\n - Any remaining tweets required to fulfill your requested sample size (e.g., 500 tweets) are loaded directly from the local database.\n - **Result**: On repeated or intraday searches, you only pay for the few brand-new tweets posted since your last search (often 1 page call = ~$0.006) instead of re-fetching hundreds of tweets you already have.\n │\n ▼\n Tier 3: TypeSafe Jev System One Evaluation (`typesafe-sdk`)\n Evaluates 4 typed questions concurrently on the combined state:\n 1. Trade Action (Choice: Strong Buy, Buy, Hold, Take Profit, Sell, Strong Sell)\n 2. Sentiment Spectrum (Score: Extreme Panic to Euphoria)\n 3. Squeeze Risk (Noul: probability that negative funding + panic indicates a short squeeze)\n 4. Catalyst Significance (Score: None, Minor, Moderate, Major)\n │\n ▼\n Decision Card Displayed in Web Terminal\n - Recommended action with calibrated confidence percentage\n - Macro sentiment gauge across all 500 tweets\n - Calculated entry range, stop loss, and target levels\n - Interactive TradingView candlestick chart" + }, + "okinaaudio/live-jev": { + "sha": "52e7d45c799f9b4d20053d2c7e28b8caac0e1b07", + "path": "README.md", + "size": 12968, + "excerpt": "# Live Jev\n**Control Ableton Live with one short sentence.**\nPress **⌘⇧Space** while you work in Live and a small bar appears on top of it. Type (or dictate) something like “turn it down 3 dB”, “Serum 2 on a new track”, or “quantize to 1/16”, then hit Enter. The bar disappears instantly, Live stays in front, and the change is applied. The bar only comes back when it needs to ask you something.\nEnglish and Japanese are both supported, with many ways to say the same thing.\n> **Status: app 1.02, Remote Script 0.18.** Live Jev is distributed as source, and this repository does not plan a packaged release.\n## What it can do\n- **Mixer** — volume, pan, mute, solo, arm, monitoring, sends (“mute”, “down by 3 dB”, “pan left 20”, “send A up a bit”). Return tracks can be addressed by letter or number and support volume, pan, mute, and solo. Master supports volume. Master cannot be muted, soloed, armed, or renamed. Returns cannot be armed. If you don’t name a target, Live Jev acts on the selected track.\n- **Transport** — play, stop, record, loop, metronome, tempo, jump to bar, Live Jev undo, capture MIDI.\n- **Clips and scenes** — launch and stop, loop, warp, pitch, gain.\n- **Notes** — quantize (1/4 to 1/32, triplets, strength), legato, transpose by octaves or semitones, velocity, double the loop.\n- **Devices** — insert plug-ins and Live’s own devices (“new track with Omnisphere”, “add EQ Eight”, “put Valhalla on the master”, “add Echo to return A”), turn devices on and off, and change parameters. Devices on returns addressed by letter or number, and devices on Master, are targets too. Instruments cannot be inserted on a return or Master. Candidates come from *your* Live browser. New tracks land where Live would put them and get Live’s default names.\n- **Tracks** — add MIDI, audio, and return tracks, and rename ordinary or return tracks.\n- **Several tracks at once** — mute, solo or arm a range, everything, or everything but one (“mute tracks 3 to 6”, “unsolo all”, “mute everything except Drums”, “solo only Bass”). These forms apply only to ordinary tracks. A single command that names two tracks, such as “mute Pad and Bass”, is not executed. Say “mute Pad and mute Bass”, or use a range, all, or except form.\n- **Chains** — up to four commands in one sentence, in order (“mute Pad and lower Bass by 3 dB”, “mute Pad, then solo Drums and arm Bass”). Every part is checked first; if one part is unclear, nothing runs. If a later part fails, the earlier ones are put back. Works in Japanese too.\n- **Undo** — Live Jev can undo one step only when it kept a receipt for the change. Receipts cover value changes such as volume, pan, mute, solo, arm, monitoring, sends, device on/off, parameters, tempo, clip settings, song switches, jumps, and renames. Live Jev writes back the value it read just before the change. It refuses if that value has since changed by hand. Inserted devices, added tracks, and note edits need Live’s own ⌘Z.\n## How it works\n- Meaning is decided by **Jev**, a small, fast model from [TypeSafe](https://typesafe.ai) that picks from a list of choices. Fixed phrases are answered locally without calling Jev at all. By default, no LLM is involved. You can enable the optional Gemini path described below.\n- Live is driven through one small Python **Remote Script** (`LiveJev`) that runs inside Live. No Max for Live required. A command takes about 20 ms; reading the whole set takes about 10 ms.\n- The bar is Swift (AppKit). The background service is Python 3.13, standard library only.\n## Getting started\n**Before you start**\n- Live Jev needs **your own API key for TypeSafe’s Jev model**. Sign in at to create one; the docs are at . It is a paid API — TypeSafe’s site listed $42 per billion input tokens in September 2026 (one command is a few hundred tokens). Check their site for current pricing.\n- Live needs a one-time setup: a small **Remote Script** is copied into Live’s User Library and selected in Live’s settings. Nothing needs to be added to your sets.\n- Live Jev is distributed as source. You build the app on your own Mac with a few Terminal commands, so you can read and change everything it does.\n### Requirements\n- Apple Silicon Mac (M1 or later), macOS 14 or later, Ableton Live 12 (Suite not required)\n- [Homebrew](https://brew.sh) and the Xcode Command Line Tools (`xcode-select --install`)\n- A TypeSafe API key\n**The step-by-step guide, with a check for every step, is in [INSTALL.md](INSTALL.md).** You can also hand that file to an AI coding assistant and ask it to set things up for you.\n### Quick start\n```bash\nbrew install python@3.13\ngit clone https://github.com/okinaaudio/live-jev.git ~/live-jev\ncd ~/live-jev\n# the Remote Script that runs inside Live\nmkdir -p ~/Music/Ableton/User\\ Library/Remote\\ Scripts/LiveJev\ncp remote_script/LiveJev/*.py ~/Music/Ableton/User\\ Library/Remote\\ Scripts/LiveJev/\n# build the app → ~/Applications/Live Jev.app\nbash scripts/build-app.sh\nopen ~/Applications/Live\\ Jev.app\n```\n**Do not move or delete the cloned folder after building** — the app runs `daemon.py` from it.\nThen, in Live: Settings → **Link, Tempo & MIDI** → **Control Surface** → choose **LiveJev** in a free slot → restart Live. Open `~/Applications/Live Jev.app` (a waveform icon appears in the menu bar; there is no Dock icon), bring Live to the front, and press **⌘⇧Space**.\nThe first launch opens a **Setup** window. Paste your TypeSafe API key there to store it in the macOS Keychain. Setup also checks the Remote Script and the connection to Live. The menu bar icon lets you reopen Setup, switch the language (Automatic / Japanese / English), and launch at login.\nFor terminal use, you can put the TypeSafe key in `~/.zshrc` instead:" + }, + "heyjunpenn/awesome-jev": { + "sha": "0b24293bc84ff0cc828b67c7d5049b74d0d085ed", + "path": "README.md", + "size": 104955, + "excerpt": "

\n \"Awesome\n

\n

\n \"485\n \"10\n \"25\n \"License:\n

\n

\n English · 简体中文 · 日本語 · 한국어 · Español · Português (Brasil)\n

\n

\n 🌐 Explore the website\n  · \n ➕ Submit a project\n

\n## About Awesome Jev\nAwesome Jev is an independent, community-maintained catalog of **485 open-source projects** built with [Jev](https://typesafe.ai/), TypeSafe AI's System One model for typed decisions inside software. It is not affiliated with or endorsed by TypeSafe AI.\n> **What makes this catalog useful?**\n>\n> Every entry identifies the concrete decision Jev makes and links to the strongest public evidence available—so you can evaluate real implementations, not just project claims.\n## Contents\n- ✅ [Official](#official-6) — **6**\n- 📦 [SDKs & clients](#sdks--clients-42) — **42**\n- 🧩 [Frameworks & integrations](#frameworks--integrations-28) — **28**\n- 🤖 [Agent tooling](#agent-tooling-112) — **112**\n- 🖥️ [Browser & computer use](#browser--computer-use-41) — **41**\n- 🪟 [Applications](#applications-53) — **53**\n- 🎮 [Games & simulations](#games--simulations-53) — **53**\n- 🧪 [Demos & playgrounds](#demos--playgrounds-46) — **46**\n- 📊 [Benchmarks & research](#benchmarks--research-92) — **92**\n- 📚 [Other lists](#other-lists-12) — **12**\nThis README is a dated snapshot of **485 unique public GitHub repositories**. Stars were captured on **2026-09-18–20** for discovery, not ranking; verify current behavior, activity, and licensing upstream.\n## Added today\n
\n52 projects added on September 20, 2026\n- **Official (2):** [TypeSafe Daggerverse](https://github.com/typesafe-ai/daggerverse), [TypeSafe Overwatch](https://github.com/typesafe-ai/Overwatch)\n- **SDKs & clients (1):** [jev-cli](https://github.com/lhotwll217/jev-cli)\n- **Frameworks & integrations (4):** [jevql](https://github.com/kylemclaren/jevql), [ground-zero](https://github.com/zavocc/ground-zero), [hiep-paseo-plugin](https://github.com/HiepPP/hiep-paseo-plugin), [jev-connector](https://github.com/adhamelhayek-lab/jev-connector)\n- **Agent tooling (13):** [tenet](https://github.com/zoidsh/tenet), [jev-belay](https://github.com/valentynkit/jev-belay), [jev-commit](https://github.com/valentynkit/jev-commit), [pi-fast-jev-compaction](https://github.com/QuentinDanblon/pi-fast-jev-compaction), [jev-flash-router](https://github.com/Ravinder82/jev-flash-router), [pi-jev](https://github.com/iefnaf/pi-jev), [pi-jev-helm](https://github.com/Z761293629/pi-jev-helm), [stepwarden](https://github.com/getexcited/stepwarden), [AskJev-MCP](https://github.com/cbruyndoncx/AskJev-MCP), [jev-compaction](https://github.com/picaye/jev-compaction), [jev-plugins](https://github.com/Pinutss/jev-plugins), [jevkeep](https://github.com/hatt-io/jevkeep), [pi-jev-router](https://github.com/gloridifice/pi-jev-router)\n- **Browser & computer use (6):** [jev-social](https://github.com/socai-io/jev-social), [jev-skip](https://github.com/valentynkit/jev-skip), [jevarena](https://github.com/raihankhan-rk/jevarena), [jev-browser-pilot](https://github.com/aidil2105/jev-browser-pilot), [jevlens](https://github.com/knowlet/jevlens), [jev-orb](https://github.com/bottlebrushes/jev-orb)\n- **Applications (7):** [jev.nvim](https://github.com/valentynkit/jev.nvim), [github-star-organizer-jev](https://github.com/yutkat/github-star-organizer-jev), [jev](https://github.com/haibt163/jev), [JevSysUno](https://github.com/Dujaydis/JevSysUno), [jev-trader](https://github.com/renatosousa/jev-trader), [newsscore](https://github.com/mahynotch/newsscore), [trading-bot-jev](https://github.com/Spykoninho/trading-bot-jev)\n- **Games & simulations (8):** [jev-plays-pokemon-red](https://github.com/valentynkit/jev-plays-pokemon-red), [jev-royal](https://github.com/Amrit-Nigam/jev-royal), [beat-jev](https://github.com/ojusave/beat-jev), [f1](https://github.com/MartinPuli/f1), [jev-atari-lab](https://github.com/memorysaver/jev-atari-lab), [jev-plays-pokemon](https://github.com/zbloss/jev-plays-pokemon), [JevArena](https://github.com/rolki-png/JevArena), [naimono-lab](https://github.com/mocchalera/naimono-lab)\n- **Demos & playgrounds (2):** [forma-system1-experiment](https://github.com/LamplighterPaul/forma-system1-experiment), [tiny-jev](https://github.com/karimatayuta/tiny-jev)" + }, + "NanmiCoder/jev-arena": { + "sha": "4eb7f2dec20a2ecdf0d74a6254f311ee8bff19a7", + "path": "README.md", + "size": 2878, + "excerpt": "

\n \"Jev\n

\n# Jev Arena\n**同一批评论，对比两个模型的速度、费用和标注结果。** 支持 CSV / Excel 导入、实时对决、录像回放和离线报告，默认 Jev vs DeepSeek，也可配置其他 OpenAI 兼容聊天模型。\n[Jev 模型原理](https://nanmicoder.github.io/jev-arena/) · [数据与报告](examples/demo/README.md) · [使用指南](docs/usage.md)\n![一万条评论历史对决的前 20 秒，1× 回放，无加速](assets/readme/demo.gif)\n[演示来源](assets/readme/demo.md) · [静态截图](assets/readme/demo-still.png)\n## 一万条评论实测\n使用 **GPT-6 Astra 对同一批 10,000 条评论全量复核**：只看原文和原标题独立标注，再对照两侧结果；另随机抽取 400 条复标。\n| 指标 | Jev 1.13 | DeepSeek Flash |\n| --- | ---: | ---: |\n| 处理耗时 | 203.2 秒 | 823.5 秒 |\n| 费用（美元） | $0.84 | $1.50（估算） |\n| 相关性准确率 | 94.70% | 96.26% |\n| 情感准确率 | 82.91% | 84.54% |\n| 意图准确率 | 77.90% | 80.33% |\n| **三项同时正确** | **62.69%** | **67.26%** |\n准确率采用允许合理歧义的口径，三项指相关性、情感、意图；**这是 AI 参考下的复核结果，不是人工金标准**。严格只认首选答案时，三项全对率为 50.58% / 55.45%。速度、费用与准确率均仅代表本次数据和配置。\n[完整准确率报告](audit/accuracy-0919-124001/report.md) · [逐条核查表](audit/accuracy-0919-124001/逐条核查.csv) · [原始数据与两侧标签](examples/demo/README.md)\n## 快速启动\n需要 **Node.js 22+**；也可使用 [Docker](docs/usage.md#使用-docker)。\n```bash\ngit clone https://github.com/NanmiCoder/jev-arena.git\ncd jev-arena\nnpm ci\nnpm start\n```\n打开 [localhost:5173](http://localhost:5173)，填入 **OpenRouter Key 和 DeepSeek Key**，保存后点击「开始对决」。默认使用 **20 条合成评论**；可上传自己的 [CSV](examples/comments.csv) / [Excel](examples/comments.xlsx)，超过 30 条需确认。\n结果保存在 `runs/<运行ID>/`，可随时回放；**回放不调用模型**。配置、导入格式与费用口径见 [使用指南](docs/usage.md)。\n## 生成报告\n运行结束后，替换为实际运行 ID：\n```bash\nnpm run report -- --run runs/<运行ID>\n```\n离线生成两侧 HTML 报告，无需 Key。默认正文为**事实草稿**；需要 Agent 撰写分析时，按 [报告生成指南](docs/report-generation.md) 操作。\n---\n[使用指南](docs/usage.md) · [标签契约](docs/CONTRACT.md) · [反馈问题](https://github.com/NanmiCoder/jev-arena/issues) · 测试：`npm test`\n[MIT License](LICENSE)。演示评论归原作者与平台，不属于代码许可范围。" + }, + "yzfly/awesome-jev-zh": { + "sha": "f8c549471e6ae4bbc3525408fda7c7e1d1feb626", + "path": "README.md", + "size": 106685, + "excerpt": "# Awesome Jev ZH\n[![Awesome](https://awesome.re/badge.svg)](https://awesome.re)\n[![PRs Welcome](https://img.shields.io/badge/PRs-welcome-black.svg?style=flat-square)](CONTRIBUTING.md)\n[![License](https://img.shields.io/badge/license-CC0--1.0-black.svg?style=flat-square)](LICENSE)\n[![自动收录](https://img.shields.io/badge/热门项目-每日自动收录-black.svg?style=flat-square)](#-热门项目自动榜)\n**Jev 不生成文本。** 第一次看到这句话时我以为是个缺陷，后来才发现这正是它的设计核心。\n它的用法是：给它一段 state（一封邮件、一行日志、一个工单），再给它几个带类型的问题，它在 70–500ms 内一次性答完——从你给的选项里选一个、在你给的量表上打一个分、或者给出一个 0 到 1 的概率，每个答案都附带一个置信度。输入 $0.042 / MTok，输出不计费。\n它的边界也很清楚：要写文案、要总结文章、要解释判断理由，LLM 仍然是更合适的工具。这一条后面还会出现好几次。\n那它解决了什么？整理这份列表的过程中，我越来越确信一件事：**我们今天写的很多 LLM 调用，本质上只是在做选择题。** 拼 prompt、逐 token 生成、剥掉 markdown 代码块、`json.loads`、校验 schema、失败了再重试——绕这么大一圈，只为了拿回 `\"billing\"` 这一个词。这段代码我自己写过不止一次。Jev 想省掉的就是这一圈。\n另外有两个数字想先摆出来，免得看完才发现：**「快 193 倍」来自 TypeSafe 自己的评测**，官方也标注了那是收益上限；而独立评测里，在钓鱼邮件这个具体任务上，直接问它一句只有 62.6% 的准确率，两行正则规则能到 91.8%。完整数据在 [冷静看待](#-冷静看待) 一节。\n这是 Jev 生态的中文精选列表，外加两份中文指南和一份 [图解说明](https://code.jiangshu.ai/awesome-jev-zh/)。\n非官方整理，与 TypeSafe AI 无隶属关系 · Jev 于 2026-09-15 开放 early access · 所有厂商自评数据都标注了出处\n---\n## 目录\n**入门** — [官方资源](#-官方资源) · [优质项目](#-优质项目) · [Jev 是什么](#-jev-是什么) · [体验渠道](#-体验渠道) · [上手](#-上手) · [规格与定价](#-规格与定价) · [该用与不该用](#-该用与不该用) · [中文指南](#-中文指南)\n**项目** — [热门自动榜](#-热门项目自动榜) · [SDK](#-sdk-与客户端) · [应用](#-应用) · [Demo](#-demo) · [Agent 工具](#-agent-工具) · [复现与评测](#-复现与评测)\n**资料** — [Cookbook 与模式](#-cookbook-与模式) · [文章](#-文章) · [社区](#-社区) · [冷静看待](#-冷静看待)\n---\n## 📘 官方资源\n官方文档写得相当清楚。真要弄懂这个模型，这里是最短的路径；二手解读（包括这份列表）只能算补充。\n| 资源 | 说明 |\n| :-- | :-- |\n| [TypeSafe 官网](https://typesafe.ai) | 官网、waitlist、产品介绍 |\n| [文档首页](https://docs.typesafe.ai/introduction) | 入门、原语、模式、API、SDK |\n| [Quick start](https://docs.typesafe.ai/introduction/quickstart) | 最短上手路径，下面 [上手](#-上手) 一节是它的中文版 |\n| [Playground](https://console.typesafe.ai/playground) | 浏览器里粘 state、加问题、看类型化结果 |\n| [API Keys 控制台](https://console.typesafe.ai/settings/keys) | 拿 `TYPESAFE_API_KEY` |\n| [HTTP API 参考](https://docs.typesafe.ai/api) | `POST https://api.typesafe.ai/v1/systemone` |\n| [Models](https://docs.typesafe.ai/models) | 模型 ID、价格、上下文与速率限制 |\n| [Confidence](https://docs.typesafe.ai/confidence) | 置信度的语义与用法，必读 |\n| [State 概念](https://docs.typesafe.ai/concepts/state) | 怎么组织喂进去的状态 |\n| [System One 概念](https://docs.typesafe.ai/concepts/system-one) | 这类模型到底是什么 |\n| [How to build with TypeSafe](https://docs.typesafe.ai/concepts/how-to-build-with-system-one) | 官方的架构心法 |\n| [用例地图](https://docs.typesafe.ai/concepts/use-case-map) | 官方列的适用场景全景 |\n| [AI 入门读本](https://docs.typesafe.ai/introduction/machine-learning-primer) | 给非 ML 背景工程师的铺垫 |\n| [Jev 1.13 能力毛边](https://docs.typesafe.ai/model-jaggedness/jev-1.13) | 官方公布的已知失败模式，上生产前必读 |\n| [Workflow evals](https://evals.typesafe.ai) | 官方公开的评测方法与结果，属厂商自评 |\n| [Agent skill 文档](https://docs.typesafe.ai/agent-skill) | 给 Claude Code / Codex 等编程 Agent 的技能包 |\n| [GitHub 组织 `typesafe-ai`](https://github.com/typesafe-ai) | 官方开源仓库 |\n| [法务条款](https://docs.typesafe.ai/legal) | 数据使用与合规 |\n官方博文，四篇立场文章：\n| 博文 | 内容 |\n| :-- | :-- |\n| [Introducing System One Models and Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 发布博文。架构、RLCD、定价、Doom 与 Wikiracing demo、FAQ |\n| [The Bitterest Lesson](https://typesafe.ai/blog/bitterest-lesson) | 它的核心论点：优化错了任务，规模再大也盖不过去 |" + }, + "openroboto-ai/jev-robot-control": { + "sha": "0e749c38197e78c46dacbb1f61910ef2b244a823", + "path": "README.md", + "size": 8306, + "excerpt": "# Jev robot control\n**Jev vs GPT-6 Astra vs GPT-4.1 mini: direct Cartesian control of an xArm7 in MuJoCo.**\nOne apple. One plate. Each controller chooses an intent, then X/Y/Z movement\ndirections and a gripper command. The shared executor applies one small motion\nincrement and returns physical feedback. This repository includes the code,\noriginal recorded responses and trajectories, an offline verifier, and a\nsynchronized three-column replay.\n[Watch/download the comparison](media/jev-vs-gpt6-vs-mini-xyz.mp4)\n![Final comparison](media/final.png)\n## Recorded result\n| Controller | Outcome | Cycles | API calls | API cost | Wall time | Simulation time |\n|---|---|---:|---:|---:|---:|---:|\n| Jev 1.13 | Placed | 113 | 226 | $0.018825 | 181.847 s | 36.16 s |\n| GPT-6 Astra, low reasoning | Placed | 106 | 212 | $5.933624 | 707.274 s | 33.92 s |\n| GPT-4.1 mini | 160-cycle limit reached | 160 | 320 | $0.288512 | 704.253 s | 51.20 s |\nThese are **one seed-0 trial per controller**, not success-rate estimates.\nThe Jev/GPT-6 recordings are the latest paired run; mini is from the earlier\npaired run with the same initial scene, physics, action instructions and cycle\nbudget. See [the protocol and source IDs](docs/RESULTS.md).\nIn these recordings, Jev's API cost was about 1/315 of GPT-6's and its wall time\nabout 26%; GPT-6 used seven fewer control cycles. Model availability, provider\npricing, network timing and fresh responses can change future results.\n## 1. Replay the published recordings\nOnly Python 3.12 and the files in this repository are needed for the recorded web\nreplay. No API key or GPU is required:\n```bash\ngit clone https://github.com/openroboto-ai/jev-robot-control.git\ncd jev-robot-control\npython incremental_triple_app.py --port 8773\n```\nOpen . Click **Play** or **Replay from start**. The page\naligns simulation time and holds each completed controller's actual final frame.\nCosts and wall times update from recorded responses. API waiting is excluded\nfrom playback and included in the wall-time metric. The server binds to loopback\nand offers read-only replay; it makes no model calls.\n## 2. Verify the physical trajectories offline\nCreate a Python 3.12 virtual environment and install the pinned dependencies:\n```bash\npython -m venv .venv\n# Linux/macOS:\nsource .venv/bin/activate\n# Windows PowerShell: .venv\\Scripts\\Activate.ps1\npython -m pip install -r requirements.txt\npython check_manifest.py\npython -m unittest discover -s . -p test_incremental.py -v" + }, + "win4r/jev-skill-suggester": { + "sha": "2a03bec83f3c9915d7c8691523a7a489c65fee7e", + "path": "README.md", + "size": 11404, + "excerpt": "# Jev Skill 建议器\n[English](README.en.md) · [真实案例](docs/examples.md) · [验证记录](docs/validation.md) · [设计说明](references/design.md)\n用 TypeSafe Jev 为当前任务推荐一个合适的已安装 Skill。先阅读技能描述筛选，再核对候选正文片段；允许返回“无需技能”或“不确定”。用户明确指定的技能通过本地查找优先处理。\n适合已安装很多技能、相邻技能用途容易混淆、需要明确决定先读哪个 `SKILL.md` 的场景。它是一个 **Codex Skill + 独立 Python CLI**，不执行或安装候选技能，不修改 Agent 设置，不添加自动运行 hook。\n## 快速开始\n要求 **Python 3.10+**。运行时仅用标准库，无需 `pip install`。离线模式不需要 Key；Jev 模式需要 [TypeSafe](https://typesafe.ai/) API Key，固定使用 `jev-1.13.0`。\n```bash\ngit clone https://github.com/win4r/jev-skill-suggester.git\ncd jev-skill-suggester\n# 读取本地技能目录，不访问 API\npython3 -I -B scripts/suggest.py catalog\n# 默认 local 模式：只给关键词候选，不给语义推荐\npython3 -I -B scripts/suggest.py suggest --task '为已有文章做公众号排版'\n# Jev 模式：隐藏输入 Key，并写入一个新的报告文件\nmkdir -p results\npython3 -I -B scripts/suggest.py suggest \\\n --task '把已有文章排成微信公众号 HTML，保留原文，不发布' \\\n --mode jev --prompt-key --out results/wechat.json\n```\n这些命令使用你自己的技能目录。本仓库不附带案例中被推荐的技能，未安装时不能得到同样结果。`local` 模式始终是 `local_only`，且 `suggestion` 为 `null`；不能把关键词第一名当作 Jev 的判断。\n### 安装为 Codex Skill\n在克隆目录中运行以下命令，只复制六个运行时文件。目标目录已存在时会报错，避免覆盖已有安装：\n```bash\npython3 - <<'PY'\nfrom pathlib import Path\nimport shutil\ntarget = Path.home() / '.codex/skills/jev-skill-suggester'\ntarget.mkdir(parents=True, exist_ok=False)\nfor name in ['SKILL.md', 'LICENSE']:\n shutil.copy2(name, target / name)\nfor name in ['agents', 'scripts', 'references']:\n shutil.copytree(name, target / name)\nprint(target)\nPY\n```\n打开新会话后调用：\n> 使用 $jev-skill-suggester，为“把已有文章排成微信公众号 HTML，保留原文，不发布”推荐合适的已安装技能。\n其他 Agent 可以通过 CLI 调用；使用 `--root` 显式指定该 Agent 的技能目录。此版本验证了 Codex Skill 与 CLI，未验证其他宿主的自动发现和调用约定。\n## 工作方式\n1. 在指定范围读取 `SKILL.md` 的名称、描述和正文，排除自身及不允许隐式调用的技能。\n2. 用户明确点名时用 `--require` 本地查找，零 API 调用。\n3. Jev 用 `Choice` 排序描述，并用单独的 `Noul` 判断该组技能是否有用；每组最多保留三个候选。\n4. 复核候选的正文片段，再用 `Choice` 与每个候选的适配 `Noul` 决定是否推荐。\n5. 宿主阅读被推荐技能的完整入口，核实当前会话可用性、用户限制和所需工具，再决定如何使用。\n复核推荐门槛为适配值 ≥ 0.80、Choice confidence ≥ 0.65。这些是探索性门槛，**不是经过校准的正确率**。Jev 提供类型化判断，不撰写推荐理由；理由应由宿主依据实际技能描述说明。" + }, + "PyModel/typesafe-mcp": { + "sha": "c7c36760aa1cfb1c4db3bafdc867b4bafd1dd241", + "path": "README.md", + "size": 15857, + "excerpt": "
\n\"typesafe-mcp\"\n`evaluate` is a stdio MCP server for [TypeSafe](https://typesafe.ai) Jev. Coding agents send observed state plus typed questions and get back probabilities, choices, and scores they can branch on in code.\nJev is a decision primitive. The host still reasons, edits, and executes.\n[![Latest release](https://img.shields.io/github/v/release/PyModel/typesafe-mcp?sort=semver)](https://github.com/PyModel/typesafe-mcp/releases/latest)\n[![GitHub stars](https://img.shields.io/github/stars/PyModel/typesafe-mcp)](https://github.com/PyModel/typesafe-mcp/stargazers)\n[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)\n![Go version](https://img.shields.io/github/go-mod/go-version/PyModel/typesafe-mcp)\n[Install](#install) | [Setup](#setup) | [The tool](#the-tool) | [Runtime](#runtime) | [Quality gates](#quality-gates) | [Contributing](#contributing)\n
\n```\n┌────────────────┐ evaluate tool ┌──────────┐ POST /v1/systemone ┌──────────────┐\n│ Claude Code │ ─────────────────▶ │ evaluate │ ─────────────────────▶ │ TypeSafe API │\n│ Claude Desktop │ MCP stdio │ binary │ or /decisions │ or │\n│ Codex │ │ │ retries 429 / 529 │ OpenRouter │\n│ pi · OMP │ ◀───────────────── │ │ ◀───────────────────── │ │\n│ PyThinker · … │ typed JSON │ │ │ │\n└────────────────┘ └──────────┘ └──────────────┘\n```\nThe tool contract is canonical in this binary: description, field docs, and agent guidance live in `cmd/evaluate/tools.go`. Claude Code, Claude Desktop, and Codex register that MCP surface. Stock [pi](https://pi.dev) has no MCP client, so `evaluate setup pi` writes an extension that registers the same tool and talks MCP to this binary. Pi that already loads MCP servers ([pi-mcp-adapter](https://github.com/nicobailon/pi-mcp-adapter) and similar) registers `evaluate mcp` in that adapter's config, same as any other stdio server. Adapters own transport and registration only; they do not author Jev semantics.\n## Install\nmacOS and Linux, amd64 and arm64:\n```sh\ncurl -fsSL https://raw.githubusercontent.com/PyModel/typesafe-mcp/main/install.sh | sh\n```\nInstalls to `~/.local/bin`. Add it to `PATH` if needed: `export PATH=\"$HOME/.local/bin:$PATH\"`. With Go 1.27+:\n```sh\ngo install github.com/PyModel/typesafe-mcp/cmd/evaluate@latest\n```\n`evaluate update` upgrades in place from a checksum-verified GitHub release.\n## Setup\nGet a key at https://console.typesafe.ai/\n```sh\nTYPESAFE_API_KEY=your-key evaluate setup\n```\nThat finds the agents installed on this machine, lists them, and asks which to register with — Enter takes all of them. `--yes` skips the question; `--host omp --host opencode` names hosts outright. Installing the binary does not register it: run this once after `install.sh`.\n| Host | How it is registered |\n|---|---|\n| Claude Code, Codex | their own `mcp add` CLI (user scope) |\n| Claude Desktop | `claude_desktop_config.json` |\n| pi (stock) | the `evaluate.ts` extension, below |\n| DSH Desktop (DeepSeek Harness) | the `mcp-evaluate` entry in `$DSH_HOME/cordis.patch.yml` (`~/.dsh` by default) |\n| OpenCode | `mcp` in `~/.config/opencode/opencode.json[c]` |\n| OMP (Oh My Pi) | `~/.omp/agent/mcp.json` |\n| PyThinker | `~/.pythinker-code/mcp.json` (or `$PYTHINKER_CODE_HOME`) |" + }, + "AkashPriyadarshii/jev-seo": { + "sha": "e3290fd15adde74b7081fca7503e65e228306419", + "path": "README.md", + "size": 11529, + "excerpt": "---\ntitle: \"jev-seo: FOSS Zero-Cost SEO & GEO Search Radar\"\ndescription: \"Open-source, subscription-free alternative to Semrush and OpenSEO. Powered by TypeSafe AI Jev and local DuckDuckGo scraping.\"\ncanonical: \"https://github.com/AkashPriyadarshii/jev-seo\"\nkeywords:\n - seo\n - geo\n - generative-engine-optimization\n - typesafe-ai\n - jev\n - rust\n - mcp\n - search-radar\n---\n\n
\n

jev-seo

\n

FOSS, zero-cost SEO & GEO search radar for developers and coding agents

\n

\n \"MIT\n \"Rust\"\n \"TypeSafe\n

\n

By Akash Priyadarshi

\n

\n Why •\n Quickstart •\n Workflows •\n Architecture •\n Non-Goals •\n Ecosystem\n

\n
\n---\n## Why jev-seo?\nSemrush and Ahrefs cost upwards of $130 per month. OpenSEO still requires paid DataForSEO credit cards. Most SEO suites are bloated web dashboards filled with vanity charts.\n- **Zero subscriptions (₹0)**: Scrapes DuckDuckGo HTML and suggest endpoints directly. No credit cards, no paid API keys.\n- **TypeSafe Jev System One**: Semantic intent classification, competitive gap detection, and AI visibility (GEO) scored deterministically without conversational LLM hallucinations.\n- **Batch Directory Auditing**: Audits hundreds of markdown/HTML files in <1 second, detecting title collisions, canonical mismatches, and thin pages.\n- **2026 Schema.org Validator**: Deeply inspects JSON-LD schemas (`SoftwareApplication`, `Article`, `Organization`, `Product`) and flags deprecated schemas.\n- **Robots.txt & AI Crawler Radar**: Evaluates permissions for AI bots (`GPTBot`, `ClaudeBot`, `PerplexityBot`, `Google-Extended`, `Bytespider`)." + }, + "devtooligan/jevscan-evm": { + "sha": "075a01007b92aec41fd2a4ff52d1c69d07840028", + "path": "README.md", + "size": 8763, + "excerpt": "```\n ██╗███████╗██╗ ██╗███████╗ ██████╗ █████╗ ███╗ ██╗ ███████╗██╗ ██╗███╗ ███╗\n ██║██╔════╝██║ ██║██╔════╝██╔════╝██╔══██╗████╗ ██║ ██╔════╝██║ ██║████╗ ████║\n ██║█████╗ ██║ ██║███████╗██║ ███████║██╔██╗ ██║█████╗█████╗ ██║ ██║██╔████╔██║\n██ ██║██╔══╝ ╚██╗ ██╔╝╚════██║██║ ██╔══██║██║╚██╗██║╚════╝██╔══╝ ╚██╗ ██╔╝██║╚██╔╝██║\n╚█████╔╝███████╗ ╚████╔╝ ███████║╚██████╗██║ ██║██║ ╚████║ ███████╗ ╚████╔╝ ██║ ╚═╝ ██║\n ╚════╝ ╚══════╝ ╚═══╝ ╚══════╝ ╚═════╝╚═╝ ╚═╝╚═╝ ╚═══╝ ╚══════╝ ╚═══╝ ╚═╝ ╚═╝\n```\n**Produce a heat map of likely bugs - in seconds, for pennies - built with TypeSafe's [Jev](https://docs.typesafe.ai).**\n> **Warning:** this is a 100% vibe-coded proof of concept. I did not read one line of the code. Use at your own risk.\n## Quick start\nPython 3.11+ and a TypeSafe API key ([console.typesafe.ai](https://console.typesafe.ai)).\n### Installation\n```sh\ngit clone https://github.com/devtooligan/jevscan-evm.git && cd jevscan-evm\npip install aiohttp\ncp .env.example .env # then set TYPESAFE_API_KEY in .env\n```\n### Usage\n```sh\npython jevscan.py /path/to/repo\n```\nResults land in `out//HEATMAP.md`. Settings (which folder, which files, thresholds) live in `jevscan.toml`; pass your own overrides with `--config` — see [docs/CONFIG.md](docs/CONFIG.md).\n## What it checks\n| Layer | Questions | From |\n| ---------- | --------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |\n| General | 2 | \"Is there any exploitable bug?\" and \"Is there a bug that lets an attacker steal or lock funds?\" |\n| Categories | 14 | OWASP Smart Contract Top 10, SWC, Immunefi ([TAXONOMY.md](docs/TAXONOMY.md)) |\n| Detectors | 343 | Sourced from [Cyfrin audit-checklist](https://github.com/Cyfrin/audit-checklist) (the Solodit checklist) and [evm-cortex](https://github.com/ccashwell/evm-cortex) by Chris Cashwell |\n## Example: USSD\nThe [Sherlock USSD contest](https://github.com/sherlock-audit/2023-05-USSD) (May 2023), 8 of 12 files scanned: [`HEATMAP.md`](bench/ussd/run/HEATMAP.md).\n> **89% chance of a critical bug · 91% chance of at least one exploitable bug · hottest file: `USSDRebalancer.sol`**\nSpeed and cost: One file, 14 checks:\n- GPT-5.6 (high reasoning) 57 seconds and 3.9¢\n- Jev 0.7 seconds and costing 0.015¢ -- about 80× faster and 260× cheaper.\n| File | Crit | Any | Access | Proxy | Oracle | Econ | Reentry | Shares | Math | Sigs | Xchain | Tokens | Logic | DoS | MEV | LowLvl |\n| ----------------------------------- | ---- | --- | ------ | ----- | ------ | ---- | ------- | ------ | ---- | ---- | ------ | ------ | ----- | --- | --- | ------ |\n| `oracles/StableOracleDAI.sol` | 🟧 | 🟥 | 🟩 | 🟩 | 🟥 | 🟨 | 🟩 | 🟩 | 🟥 | 🟩 | 🟩 | 🟩 | 🟨 | 🟨 | 🟩 | 🟩 |\n| `USSDRebalancer.sol` | 🟥 | 🟥 | 🟧 | 🟨 | 🟥 | 🟨 | 🟧 | 🟥 | 🟥 | 🟩 | 🟩 | 🟥 | 🟥 | 🟥 | 🟥 | 🟩 |\n| `USSD.sol` | 🟥 | 🟥 | 🟥 | 🟧 | 🟥 | 🟧 | 🟧 | 🟧 | 🟥 | 🟩 | 🟩 | 🟥 | 🟥 | 🟧 | 🟥 | 🟩 |\n| `oracles/StableOracleWBTC.sol` | 🟧 | 🟥 | 🟩 | 🟩 | 🟥 | 🟨 | 🟩 | 🟩 | 🟧 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 |\n| `oracles/StableOracleWETH.sol` | 🟧 | 🟧 | 🟩 | 🟩 | 🟥 | 🟩 | 🟩 | 🟩 | 🟧 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 |\n| `oracles/StableOracleWBGL.sol` | 🟨 | 🟧 | 🟩 | 🟩 | 🟥 | 🟩 | 🟩 | 🟩 | 🟧 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 |\n| `oracles/UniswapV3StaticOracle.sol` | 🟨 | 🟨 | 🟩 | 🟩 | 🟨 | 🟩 | 🟩 | 🟩 | 🟨 | 🟩 | 🟩 | 🟩 | 🟨 | 🟩 | 🟩 | 🟩 |\n| `Migrations.sol` | 🟩 | 🟨 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 | 🟩 |" + }, + "mizzlelover/jev-hub": { + "sha": "04530e25b75136c90c0e65569dc13fb07a55aab1", + "path": "README.md", + "size": 55472, + "excerpt": "# JEV HUB · Jev 聚合站\n> X（Twitter）上关于 **TypeSafe AI「系统一模型」Jev** 的长文与演示视频聚合。\n> 逐条保留**原帖链接**与**作者署名**，本站不转载、不二次托管任何内容。\n**[→ 打开聚合网页浏览全部内容](https://mizzlelover.github.io/jev-hub/)** ｜ **[结构化数据 data/posts.json](data/posts.json)**\n由 **谁是专家** 出品（小红书 / 微信公众号 / X 同名）。\n## 数据概览\n| 指标 | 数值 |\n| --- | --- |\n| 动态总数 | 714 |\n| 演示视频 | 226 |\n| 长文 | 114 |\n| 讨论 | 374 |\n| 参与作者 | 604 |\n| 语言分布 | 英文 513 · 中文 122 · 日文 74 · 韩文 2 · 阿拉伯文 3 |\n| 数据跨度 | 2026-09-15 18:17 ~ 2026-09-19 02:19（UTC） |\n| 最近更新 | 2026-09-19 02:44（UTC） |\n## 🐣 小白入门（先读这篇）\n**[读不懂也要读懂系列01｜GPT它爸，造了个不会聊天的 AI](https://x.com/dboy_yi2025/status/2100888192576880692)**\n> 写给完全不懂的人看的 Jev 科普：从「一个不会说话的 AI 有什么用」讲起，把 Jev 为什么快、为什么便宜、能干什么活一次讲清。小白先读这篇，再看下面的演示和长文就不迷路了。\n— [@dboy_yi2025](https://x.com/dboy_yi2025) · 2026-09-18\n## 官方入口\n- 官网：[typesafe.ai](https://typesafe.ai)\n- OpenRouter：[openrouter.ai/typesafe/jev](https://openrouter.ai/typesafe/jev)\n- Vercel AI Gateway：[vercel.com/ai-gateway](https://vercel.com/ai-gateway)\n- Venice API：[venice.ai](https://venice.ai)\n## ⭐ 编辑精选（按互动量）\n- After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a ne… — [@CompleteSkeptic](https://x.com/CompleteSkeptic/status/2099925682726002904) · 2026-09-15 · `模型原理 · 成本速度`\n- found the perfect use case for @typesafeai Jev: instant compaction in 2026, why is compaction still a summarization prompt? Jev can make it instant by… — [@tamarajtran](https://x.com/tamarajtran/status/2100694549362553153) · 2026-09-17 · `智能体`\n- The gains aren’t free: Jev can't generate text Comparing Jev vs LLMs side-by-side makes the trade-off clear Fun fact: replacing sequential computation… — [@CompleteSkeptic](https://x.com/CompleteSkeptic/status/2099925684256899543) · 2026-09-15 · `编程开发 · 成本速度 · 评测对比`\n- Here's a 45-second TL;DR on Jev. I find the core idea beautifully simple, but the video made it really hard to understand. Hope you find it helpful. — [@MatijaSosic](https://x.com/MatijaSosic/status/2100190746389135772) · 2026-09-16 · `编程开发`\n- 【Breaking News】ChatGPT Co-Inventor Unveils New AI Model \"Jev\" After 2 Years of Stealth Development ・20-200x faster than LLMs, 40-400x cheaper ・Output … — [@ctgptlb](https://x.com/ctgptlb/status/2100120850754412967) · 2026-09-16 · `成本速度 · 评测对比`\n- jev is INSANE. in 40 seconds it broke down 724 live ads from 37 brands. every hook. every format. offer. cta. awareness stage. landing page mismatch. … — [@TheMattBerman](https://x.com/TheMattBerman/status/2100654891756589230) · 2026-09-17 · `成本速度`\n- we are officially out of stealth! join the frontier and get access to Jev on our website (link on profile) — [@typesafeai](https://x.com/typesafeai/status/2099944756931596454) · 2026-09-15\n- I created an EC real-time customer service demo using Jev and gpt-live-1. Even in the middle of a conversation, it immediately suggests recommended pr… — [@rinte0321](https://x.com/rinte0321/status/2100736454850908344) · 2026-09-17\n- New experiment: json-render + jev The future Generative UI is instant Your components, your actions, your design system Rendered in milliseconds — [@ctatedev](https://x.com/ctatedev/status/2101022101750571357) · 2026-09-18 · `编程开发 · 成本速度`\n- Jev by @typesafeai is now on OpenRouter, in beta. Jev is a System One model. Instead of generating text, it takes your app's state plus a typed questi… — [@OpenRouter](https://x.com/OpenRouter/status/2100744709589316009) · 2026-09-18 · `编程开发 · 模型原理`\n- All of the coolest Jev projects I could find on X today 🧵 — [@moritzkremb](https://x.com/moritzkremb/status/2100895894287839255) · 2026-09-18\n- Jev from @typesafeai is on AI Gateway. Build agents that decide, route, score, and stop in milliseconds: 𝚊𝚠𝚊𝚒𝚝 𝚎𝚟𝚊𝚕𝚞𝚊𝚝𝚎({ 𝚖𝚘𝚍𝚎𝚕: '𝚝𝚢𝚙𝚎𝚜𝚊𝚏𝚎-𝚊𝚒/𝚓𝚎𝚟', 𝚜𝚝… — [@vercel_dev](https://x.com/vercel_dev/status/2100378959653507175) · 2026-09-17 · `编程开发 · 智能体 · 成本速度`\n## 🎬 演示视频精选（Top 60）\n- After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a ne… — [@CompleteSkeptic](https://x.com/CompleteSkeptic/status/2099925682726002904) · 2026-09-15 · `模型原理 · 成本速度`\n- found the perfect use case for @typesafeai Jev: instant compaction in 2026, why is compaction still a summarization prompt? Jev can make it instant by… — [@tamarajtran](https://x.com/tamarajtran/status/2100694549362553153) · 2026-09-17 · `智能体`\n- Here's a 45-second TL;DR on Jev. I find the core idea beautifully simple, but the video made it really hard to understand. Hope you find it helpful. — [@MatijaSosic](https://x.com/MatijaSosic/status/2100190746389135772) · 2026-09-16 · `编程开发`\n- jev is INSANE. in 40 seconds it broke down 724 live ads from 37 brands. every hook. every format. offer. cta. awareness stage. landing page mismatch. … — [@TheMattBerman](https://x.com/TheMattBerman/status/2100654891756589230) · 2026-09-17 · `成本速度`\n- we are officially out of stealth! join the frontier and get access to Jev on our website (link on profile) — [@typesafeai](https://x.com/typesafeai/status/2099944756931596454) · 2026-09-15\n- I created an EC real-time customer service demo using Jev and gpt-live-1. Even in the middle of a conversation, it immediately suggests recommended pr… — [@rinte0321](https://x.com/rinte0321/status/2100736454850908344) · 2026-09-17" + }, + "davila7/jev-explained": { + "sha": "7c69d941dc821e2df3a37e3348357f04b0b2adb8", + "path": "README.md", + "size": 5548, + "excerpt": "# Jev Explained\n**Learn how TypeSafe's Jev makes typed, probabilistic decisions — by running it.**\nLive demo: **https://jev-explained-repo.vercel.app/** (bring your own TypeSafe or Vercel AI Gateway key).\n

\n \"Jev\n

\nAn interactive playground that shows, step by step, how [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) — TypeSafe's System One model — works.\n## What is Jev?\nJev is not a chat LLM. It does not generate text. You send it a **state** (any text or JSON: an email, a market snapshot, a tool call an agent wants to run, a whole inbox) plus one or more typed **questions**, and it returns calibrated probabilities for every question in a single ~100 ms round trip. Your code, not the model, makes the final decision by thresholding on those numbers.\n### The three primitives\nChoose the primitive by the type of question you are asking:\n| Primitive | Ask it when | Example | Returns |\n| --- | --- | --- | --- |\n| **Noul** — yes / no? | the question is binary | *Is this email spam?* | one probability, 0 → no, 1 → yes |\n| **Choice** — which one? | you pick from known options | *Which team should handle this?* | the chosen option, a probability for every option, and a `confidence` |\n| **Score** — how much / what level? | you grade on an ordered rubric | *How risky is this?* | a weighted score, a probability for every level, and a `confidence` |\nTwo things make this different from asking an LLM:\n- **Questions run in parallel.** Jev reads the state once and answers every question at the same time, so ten questions cost about the same as one. You can fan out speculatively and let your code decide what matters.\n- **Confidence is a second axis.** Choice and Score answers tell you *what* (the answer) and *how sure* (the shape of the distribution). High confidence → act automatically; low confidence → ask a human.\n### What this repo shows\nThe playground walks through four patterns, each with real requests you can run with your own key:\n| Example | Pattern | State | Questions |\n| --- | --- | --- | --- |\n| **Email Spam Classifier** | Text classification | an email | `is_spam` (noul), `folder` (choice), `suspicion` (score) |\n| **NVIDIA: Buy or Sell?** | Decision on structured data | a JSON market snapshot | `action` (choice), `sentiment` (score), `material_risk` (noul) |\n| **Agent Tool-Call Guardrail** | Jev inside an agent harness | a tool call the agent wants to run | `verdict` (choice), `is_destructive` (noul), `blast_radius` (score), `in_scope` (noul) |\n| **Inbox Triage** | Speculative fan-out | 8 support tickets | 8 × `priority` (score), `most_urgent` (choice), `needs_incident` (noul) — one request |\nFor every run the right-hand panel shows the exact request, latency and token usage, the typed answers with probability bars, and the decision your code makes from them.\nEndpoint: `POST https://api.typesafe.ai/v1/systemone` · Model: `jev-latest`. See the [API reference](https://docs.typesafe.ai/api).\n## Run it\n```bash\nnpm install\nnpm run dev\n```\nOpen http://localhost:3000, pick a provider, paste its key, choose an example and press **Run** (or ⌘↵). Edit the state or switch between the sample states to see how the answers move. Use the `≡ / ` toggle to see the raw request JSON.\n## Providers\nBoth providers speak TypeSafe's native request/response shape; only the URL, key and model id change.\n| Provider | Endpoint | Model | Key |\n| --- | --- | --- | --- |\n| TypeSafe | `https://api.typesafe.ai/v1/systemone` | `jev-latest` | [console.typesafe.ai/keys](https://console.typesafe.ai/keys) |\n| Vercel AI Gateway | `https://ai-gateway.vercel.sh/typesafe/v1/systemone` | `typesafe-ai/jev` | AI Gateway API key from your Vercel team ([docs](https://vercel.com/docs/ai-gateway/sdks-and-apis/typesafe)) |\n## How the key is handled\nNeither API accepts cross-origin browser calls, so the app ships a tiny proxy at `src/app/api/jev/route.ts`. Keys are stored per provider in your browser's `localStorage` only and forwarded on each request in the `x-jev-api-key` header (with `x-jev-provider` selecting the upstream); the server never persists them.\n## Project layout\n```" + }, + "arunav25/jev-mcp": { + "sha": "a858bf2227da3012651457b31404f70f08fb807f", + "path": "README.md", + "size": 10859, + "excerpt": "# JEV MCP · Structured Judgments & LLM Evaluation\nConnect JEV to MCP clients and compare its judgments against general-purpose LLMs using shared\ndatasets and measurable accuracy.\nAgents are good at producing text and bad at producing answers you can branch on. Ask one \"is this\nticket urgent?\" and you get back a sentence you then have to parse, with no number attached — no way\nto tell a confident yes from a coin flip. This server exposes TypeSafe's Jev classifier as a single\nMCP tool that returns typed answers with probabilities, so the agent gets `0.94` and moves on.\n```\n agent ──── evaluate ────▶ jev-mcp ──── POST /v1/systemone ────▶ TypeSafe\n (stdio) (this) (Jev)\n ◀─── typed JSON ─── ◀─── probabilities ────────\n jev-eval ── same question ─▶ JEV ─┐\n ─▶ LLM ─┴─▶ Brier · calibration · McNemar\n```\nTwo halves: an MCP server that exposes the classifier to your agents, and an evaluation harness that\ntells you whether it is actually beating whatever you were using before.\n## Requirements\n- Node.js 20 or newer\n- A TypeSafe API key from [console.typesafe.ai](https://console.typesafe.ai/)\n## Install\nNot published to npm yet, so install from the repository:\n```sh\nnpm install -g github:arunav25/jev-mcp\n```\nOr clone it, which is what you want if you plan to run the evaluation harness:\n```sh\ngit clone https://github.com/arunav25/jev-mcp.git\ncd jev-mcp && npm install && npm link\n```\n> The bare name `jev-mcp` on npm belongs to an unrelated project. This package publishes as\n> `@arunav25/jev-mcp`; until it is published, use one of the commands above.\nThen point your agents at it. The key has to be in your environment *before* you run this, because\nagents launch the server without your shell, so its value is written into each client's config:\n```sh\nexport TYPESAFE_API_KEY=sk-...\njev-mcp install\n```\nThat registers the server with Claude Code, Claude Desktop and Codex, skipping any that aren't\ninstalled. Restart Claude Desktop afterwards. To see what it would do first:\n```sh\njev-mcp install --dry-run\njev-mcp install --client codex # or limit it to one\n```\n
\nRegistering by hand" + }, + "endman100/research-Qwen3.8-JevLike": { + "sha": "7c908289d0e8a5ba36dbef64034ecf281dd2d89a", + "path": "README.md", + "size": 5354, + "excerpt": "# Qwen3.8 Jev-like：71 類二元分類\n## 前提\n本實驗起點是 HF [harshatheg/Qwen-2.5-1B-RLCD](https://huggingface.co/harshatheg/Qwen-2.5-1B-RLCD) 的 **Parallel Constrained Decoding**，另參考其 [Transformers 版本](https://huggingface.co/shreyansh26/Qwen-2.5-1B-RLCD)：\n對有限選項欄位共用前綴、平行判斷，減少逐 token 生成完整 JSON 的工作。71 類各自回答 `true/false`，可同時成立，適合用來檢驗這個思路。\n## 實驗目標\n**嘗試將上述 HF repo 的方法思路套用到 Qwen3.8-27B-NVFP4＋vLLM，確認相較直接 Structured Output 是否更快、快多少。**\n固定同一模型、任務與 71 類定義，各組測 10 次；比較取得完整 71-key bool dict 的總耗時與答案差異，並以分段計時分析瓶頸。\n本次以 vLLM prefix caching＋批次請求實作其核心思路，未直接移植原 repo 的引擎或 Tree Parallel，亦未進行 Jev／RLCD 訓練；不預設能重現原作者的加速倍率。\n## 結果\nRTX 5090 32GB、Windows 11／WSL2、vLLM 0.29.0。**同一個任務，各組 10 次**，不是 10 種任務。\n| 方法 | Context／client workers／server sequences | 計時條件 | 中位數 |\n|---|---|---|---:|\n| A：JSON Schema | 32K／1／4 | 熱前綴 | 46.589 s |\n| B：二元分類 | 32K／4／4 | 熱前綴 | 4.398 s |\n| B：四路分段 | 32K／4／4 | 冷前綴＋分類＋組裝 | 5.040 s |\n| B：71 路分段 | 6K／71／71 | 冷前綴＋分類＋組裝 | **3.465 s** |\nA/B 熱前綴快 **10.59×**，每輪 **6 類答案不同**。71 路比四路分段快 **1.45×**，另有 **3 類翻轉**。\n後兩組在不同時段測量、同時改動 context 與並行度，不能單獨歸因於某個參數。\n沒有人工標註；一致率不等於正確率，機率未校準。\n
\n全部 10 次總耗時（秒）\n| 次數 | A 熱前綴 | B 四路熱前綴 | B 四路冷分段 | B 71 路冷分段 |\n|---:|---:|---:|---:|---:|\n| 1 | 46.212 | 4.640 | 5.311 | 3.457 |\n| 2 | 47.421 | 3.979 | 4.991 | 3.454 |\n| 3 | 46.492 | 4.285 | 5.109 | 3.516 |\n| 4 | 47.272 | 4.613 | 4.810 | 3.462 |\n| 5 | 45.242 | 4.736 | 5.158 | 3.544 |\n| 6 | 47.027 | 4.122 | 5.058 | 3.508 |\n| 7 | 46.685 | 3.995 | 4.767 | 3.460 |\n| 8 | 45.767 | 4.511 | 5.402 | 3.455 |\n| 9 | 45.273 | 4.580 | 4.797 | 3.488 |\n| 10 | 46.801 | 4.043 | 5.022 | 3.468 |\n
\n## 方法與理論值\nA 逐 token 寫 JSON；B 重用共同前綴，每類只生成 **1 token**（`true=1802`／`false=3721`），再組成 dict。\n兩者均關閉 thinking、temperature=0。B 將兩候選 logprobs 正規化得到 P(true)，以 **0.5** 判定。\n分段測試每輪使用新 cache salt：冷前綴 → 71 類 → 組裝，暖機不計。前綴請求多產生一個丟棄 token，\n並非純 GPU prefill。總時間包含 HTTP、排隊與中途讀取指標，不包含載入模型及存檔。\n最新平均：**0.704 s 前綴＋2.771 s 分類＋0.000649 s 組裝**，另約 0.005 s 指標開銷。\n**完全平行的理想情境：** `T = T_prefix + max(T₁…T₇₁) + T_assembly`。\n假設 71 類都能維持舊四路平均 0.2406 s、完全重疊且無資源競爭：\n`0.7042 + 0.2406 + 0.000649 ≈ 0.9455 s`，含量測開銷約 **0.95 s**。\n這以平均代替未知的最慢分支，**不是實測或硬體下限**。71 個 HTTP 並行也不等於一次 GPU forward。\n## 重現" + }, + "marcosmartinez/jev-acento": { + "sha": "994943244b6ba1fdaf3420b2dcf5a9066f158509", + "path": "README.md", + "size": 12639, + "excerpt": "# jev-acento\n**Does Jev understand your accent?**\nAn independent, reproducible audit of [Jev](https://typesafe.ai) — TypeSafe AI's \"System One\"\nevaluation model — on **Spanish**, plus a CLI that lets anyone run the same comparison on their\nown labelled data.\n**Run `20260921-es-v1`** — 19,200 calls, 3,200 paired items, model `jev-1.13.0` (version-pinned),\nUSD 0.58, 0 errors. Every number below comes from [`results.json`](results.json) via\n`make reproduce`; none is typed by hand.\n## Findings\n**1. Spanish costs accuracy on every dataset.** Holding the instructions in English and swapping\nonly the `state` from English to Spanish (B − A), Jev is measurably worse on all four:\n| Dataset | A (EN state) | B (ES state) | Δ accuracy | Verdict |\n|---|---|---|---|---|\n| XNLI | 0.850 | 0.786 | −6.4 pp `[−8.6, −4.3]` | **measurably worse** |\n| PAWS-X | 0.834 | 0.772 | −6.2 pp `[−8.5, −3.6]` | **measurably worse** |\n| MASSIVE | 0.845 | 0.808 | −3.7 pp `[−5.8, −1.5]` | **measurably worse** |\n| Belebele | 0.982 | 0.952 | −3.0 pp `[−4.5, −1.7]` | **measurably worse** |\n**2. It also costs calibration, on the two hardest tasks.** ECE roughly doubles on XNLI\n(0.057 → 0.101) and PAWS-X (0.033 → 0.078) — *less calibrated* under the pre-registered rule.\nOn MASSIVE and Belebele the change is not detectable. Calibration matters more than accuracy\nhere: the operational consequence is that automating at `p_max ≥ 0.9` covers **72.2% of XNLI in\nEnglish but only 63.4% in Spanish**, and the items you do automate are *less* accurate\n(0.938 → 0.904), not more.\n**3. Writing the instructions in Spanish does not help.** This is the question nobody had\nmeasured, and the answer is a clean null on three of four datasets (C − B):\n| Dataset | Δ accuracy | Verdict |\n|---|---|---|\n| XNLI | −0.2 pp `[−1.1, +0.7]` | no detectable difference |\n| MASSIVE | −0.7 pp `[−1.8, +0.7]` | no detectable difference |\n| Belebele | +0.5 pp `[+0.0, +1.2]` | no detectable difference |\n| PAWS-X | +1.6 pp `[+0.7, +2.6]` | ambiguous — real but below the 3 pp threshold |\nCalibration shows no detectable difference on all four. **Practical advice: keep your\n`instructions` and `criteria` in English.** It is never worse, it is what the model is\ndocumented to be best at, and on MASSIVE the Spanish wording costs 6.2% more input tokens for\nnothing.\n**4. Spanish text costs 17–38% more input tokens** than the same content in English\n(state-only, with the fixed question overhead subtracted). Far less than the ~3× the Russian\naudit found for Cyrillic.\n### Compared with the Russian audit\n| | Russian (prior work) | Spanish (this repo) |\n|---|---|---|\n| XNLI accuracy | 88.3% → 77.3% (−11.0 pp) | 85.0% → 78.6% (−6.4 pp) |\n| XNLI ECE | 0.032 → 0.096 | 0.057 → 0.101 |\n| Token ratio | ~3× | ~1.23× |\nSpanish degrades less than Russian, which is what you would expect from a Latin-script language" + }, + "joelakaufmann-lgtm/NRS-Navigator": { + "sha": "dbe9da1004bdce5116f5aa9cee46ead9b40e37bd", + "path": "README.md", + "size": 17445, + "excerpt": "# NRS Navigator\n**Find Nevada statutes quickly—and measure whether Jev makes the search results more relevant.**\nNRS Navigator is a local research application created by **Joel Kaufmann**. It combines exact NRS citation lookup, fast keyword search, a source-text reader, and optional Jev-assisted reranking. Researchers can inspect statutory text, follow references, and save provisions or passages with their source provenance.\nThe project also serves as a practical test bed for **TypeSafe's Jev**: does reranking a fixed keyword shortlist help a researcher find the relevant statute sooner? It preserves keyword and assisted results, model judgments, failures, and evaluation artifacts so that improvements can be measured rather than assumed. Jev ranks source passages; the app does not generate legal answers.\n## Public testing site\nThe GitHub Pages edition provides browser-only full-text keyword search and reading across the recorded 835-chapter snapshot, plus the recorded Jev benchmark reports and blinded reviewer workbook. It does not run the Python backend, accept API keys, make live Jev calls, or store research collections. Its caption-weighted browser ranking is a separate implementation from SQLite FTS5; the displayed Jev comparisons are the preserved backend benchmark results.\nSearch queries remain in the browser; public catalog/index and selected chapter files are fetched from GitHub Pages. The local app remains the place to test live Jev reranking and private collections.\n## Two purposes\n1. **Quick NRS research.** Enter a citation such as `NRS 163.003` or a question such as “removal of trustee for breach of trust.” Read the stored statute, follow explicit links, and collect source-grounded research locally.\n2. **Inspectable Jev evaluation.** Compare keyword order with Jev on the same candidates; measure known-target Hit@5 and reciprocal rank; use blinded independent labels for precision, pooled nDCG, and relevance-probability calibration. Missing candidates and provider fallbacks remain visible.\nThis is a working development research tool and an independent evaluation project. It is not an official Nevada Legislature or TypeSafe product. The implementation follows [the project plan](NRS-Navigator-Full-Stack-Plan.md).\nCopyright **2026 Joel Kaufmann**. Licensed under **[Apache License 2.0](LICENSE)**. See [NOTICE](NOTICE) and [third-party materials](THIRD_PARTY_NOTICES.md).\n![Research workspace](docs/screenshots/full-corpus-research.png)\n## Start the app\nPrerequisites: Python 3.12+ and Node 22.12+ (tested with Python 3.14.6, Node 22.22.2, macOS 27.0).\n```sh\npython3 -m venv .venv\n.venv/bin/python -m pip install -r requirements.lock\nnpm --prefix apps/web ci\nnpm --prefix apps/web run build\n.venv/bin/python scripts/import_pilot.py\n.venv/bin/python scripts/launch.py\n```\nThe launcher opens the bundled app at `127.0.0.1:8765`. It prints a private launch URL containing an ephemeral token in its fragment. Use that URL after restarting the server; a plain URL in a new tab is not authorized. The token stays in tab session storage and never goes to the official source site. Stop with Control-C.\nAfter setup, local search, browsing, collections, notes, and export require **no internet connection**. Optional Jev ranking sends an explicitly approved query and bounded source passages to TypeSafe. Opening an official-source link is an explicit external navigation. There are no remote fonts or telemetry. Do not expose this development server to a LAN or reverse proxy.\n## What works\n- A full-corpus acquisition workflow, validated in the recorded run against 835 official-index chapter pages and 49,746 provision records. A 12-chapter fixture is included for quick offline setup and tests.\n- Strict contents/body reconciliation, history notes, subsection paragraph offsets, and explicit statutory links.\n- Exact citation lookup (including preserved subsection requests), caption-weighted SQLite FTS5 search, and a full section reader.\n- Multiple local research collections, full-provision or selected-passage saves, notes, ordering, and deletion.\n- Pinned source versions that survive source changes; failed imports leave the last valid version active.\n- Markdown export preview with notes excluded by default, plus a JSON source manifest. Quotations are copied from saved source spans.\n- Corpus coverage, revision/retrieval distinctions, quarantined import reports, and missing-corpus states.\n- Jev relevance ranking, typed relationship labels, original-order comparison, preview-bound approval, cancellation, caching, usage records, and keyword fallback.\n- Loopback serving, per-launch API token, host/origin checks, request limits, CSP, and plain-text rendering.\nCreate a collection, then use Research to find `NRS 163.003`. Select text in the reader to save a passage, or save the whole provision. In Collections, save any note edits before previewing an export. Export includes saved notes only when the checkbox is selected.\n## Corpus scope and limitations\nThe September 20, 2026 acquisition covered **all 835 chapter pages linked by its preserved official NRS index**, including the preliminary chapter and letter-suffixed chapters: **49,746 provision records and 354,124 body/history paragraphs and chapter notices**. All acquired pages passed source-hash, chapter-identity, contents/body and independent paragraph-order checks. Every extracted explicit NRS section link resolves within this snapshot. See [verification](docs/full-corpus-verification.json), [acquisition metadata](docs/full-corpus-acquisition.json), and [full-corpus operations](docs/full-corpus.md). The 121.9 MB original HTML download and working databases are not included in this source repository. The Pages edition includes a generated public-only normalized text corpus and search index in `docs/data/`. Recorded counts describe that snapshot, not a guarantee about future downloads.\nThe quick-start commands import the included 12-chapter fixture. No TypeSafe key is required for local search. To acquire and install the full corpus into a separate project-local workspace:\n```sh\n.venv/bin/python scripts/full_corpus.py fetch\n.venv/bin/python scripts/full_corpus.py verify\n.venv/bin/python scripts/full_corpus.py install --target .local/full-workspace\nNRS_DATA_DIR=.local/full-workspace .venv/bin/python scripts/launch.py\n```" + }, + "Rizzo-AI-Academy/rizzo-flow": { + "sha": "529432bebd5f6692f1ccffe6bbb0176dbc446aa9", + "path": "README.md", + "size": 23362, + "excerpt": "
\n\"Rizzo\n# Rizzo Flow\n### The open, local take on Jev: typed decisions from an LLM, without generating a single token\n**_Unstructured state in → typed, probabilistic decisions out. On your own machine._**\n

\n\"100%\n\"0\n\"Jev-compatible\n

\n

\n\"Spark-X2.5\n\"1M-token\n\"MLX\n\"about\n\"about\n\"Apache-2.0\n

\n🌐 Website · A project by Rizzo AI Academy · 🇮🇹 Documentazione dettagliata in italiano\n
\n**Rizzo Flow** is an open-source, local-first implementation of the idea behind\n[**Jev**](https://typesafe.ai/blog/introducing-system-one-models-and-jev), TypeSafe's \"System One\"\nmodel: a *function call with judgment* that takes unstructured state and returns **typed decisions\nwith probabilities** — a yes/no, a choice among options, a score on a rubric, a number — instead\nof text you then have to parse.\nJev is a closed, hosted service. Rizzo Flow gives you the same programming model **on your own\nhardware, with open weights, and with the same HTTP interface**, so code written against the\nTypeSafe API can point at `localhost` by changing one URL.\n> **Independent project.** Rizzo Flow is not affiliated with TypeSafe and does not reproduce Jev's\n> proprietary architecture or its RLCD training. It reproduces the *interface pattern* with an\n> off-the-shelf open model, in the spirit of [SemIf](https://github.com/TheoLeeCJ/SemIf), which\n> inspired it. Probabilities are **uncalibrated** unless you calibrate them on your own data, and\n> we make no claim of matching Jev or SemIf in quality. Every number below comes with its caveats.\n
\n
\n\"The\n🦔 The built-in playground — one support ticket, two questions answered in parallel in 484 ms:\none state prefill (116 tokens, 187 ms), one micro-batch, 0 generated tokens.
\nSpark-X2.5-4B at 8 bit on an M4 Pro · interface in Italian or English · run it yourself ↓\n
\n---\n## Why \"System One\"\nAn LLM asked to classify something *writes* an answer: token by token, slowly, in a format you\nhope is valid JSON. But for a decision you do not need text — you need **which option, and how\nsure**. That information is already in the model after a single forward pass: it is the" + }, + "1816586742-stack/jev-craft": { + "sha": "b614d4f0b3133685b56d87d2eab4570abdedb92e", + "path": "README.md", + "size": 13263, + "excerpt": "# jev-craft ｜ 盗天是也 · 卷一「反射弧」\n> **先说清这份东西是谁的**：这是「盗天是也」这个世界的**零件档案**，不是通用 AI 技巧仓库。\n> 世界宪章在 [`docs/世界宪章.md`](docs/世界宪章.md)（先读它，再读技术）。\n> 本仓库 = 卷一：**让这个世界的居民，在\"值得判断的一瞬间\"用几毫秒拿到可信答案。**\n> **一句话（对外的说法）**：Agent 慢，慢在它每做一个决定都要**写一段话再解析**。\n> 把那些\"答案空间事先能定义\"的决定交给 System One 模型（Jev），\n> Agent 就剩两种动作——**写**（交回心智）和**选**（眨眼之间出结果）。\n>\n> **一句话（对自己说的）**：天把时间扣在它那边，这一卷把它拿回来。\n[![status](https://img.shields.io/badge/impl-75%20assertions%20passing-brightgreen)](#-先验证再采信)\n[![license](https://img.shields.io/badge/license-MIT-blue)](LICENSE)\n---\n## 0. 在这个世界里，它是哪个器官\n| 世界里的器官 | 技术层 | 位置 |\n| --- | --- | --- |\n| **底线**（本能，来不及想） | 反射层：纯代码硬编码 | 最高优先，命中即抢占 |\n| **判断**（这一拍做什么） | **System One 判断层 —— 本仓库** | 高频热路径 |\n| **心智**（想清楚 / 写东西 / 记事） | 慢思考主脑 + 记忆树 | 只在升级、人类插话、目标完成时介入 |\n| **眼**（看见世界） | 多模态模型 + 结构化采集 → 全英文 state | 与判断并发，异步旁路 |\n顺序不能错：**底线 > 判断 > 心智**。反过来就不是这个世界的居民，是一个客服机器人。\n---\n## 1. 这个思路在解决什么\n现在的 Agent 循环里，大量决定**不需要一段话**：\n- 这一步该挖、该走、该撤，还是该问人？\n- 眼下这步风险大不大？值不值得做？\n- 这条路堵了，原因更可能是\"没工具\"还是\"没材料\"？\n- 这句话该不该现在说给用户听？\n生成式模型处理这类问题要付出三层代价（官方口径 + 实测都指向同一件事）：\n| 代价 | 表现 |\n| --- | --- |\n| 慢 | 想要一个 0–2 的分数，它先写 372 token 的推理再给你数字；实测延迟 1.6–3.7s |\n| 贵 | 你在为\"它向你解释自己\"付费 |\n| **脆** | 提示词一层约定 + 解析一层约定，任何一层漂移就**静默失败**（大小写、拼错、带引号，直接写进你的库） |\nSystem One 模型（Jev）把这条路彻底换掉：**不生成文字，只返回带类型的答案和校准过的概率**。\n```jsonc\n// 同一个 state，一次请求，一次拿回全部判断（答案空间是你给的，它只能在里面选）\n{\n \"next_action\": { \"choice\": \"flee\", \"probabilities\": {\"flee\": 0.71, \"dig\": 0.22, \"other\": 0.07}, \"confidence\": 0.42 },\n \"risk_of_step\": { \"score\": 2.4, \"confidence\": 0.81 },\n \"needs_light\": { \"noul\": 0.93 },\n \"escalate\": { \"noul\": 0.18 }\n}\n```\n**它不解释，但概率告诉你\"敢不敢动\"。**\n---" + }, + "Xubqpanda/JevLoop": { + "sha": "0e3f67c72801630bee7149c7e8157ea1370c37fa", + "path": "README.md", + "size": 11730, + "excerpt": "# JevLoop\n**The agent loop where decisions don't cost a model call.**\nEvery fork in a normal agent loop — *should I act? which tool? is this safe? did it work? am I done? can I ship this?* — is answered by a full LLM call. None of those are generation. They're picks, scores and yes/no answers.\nJevLoop routes them to a decision model ([Jev](https://typesafe.ai) / [Laya](https://github.com/NandaKishorM/laya)) and keeps the LLM for the one thing only it can do: **writing**.\n```\n$ npm run demo # 全新 clone，无 key、无网络\n 判定后端 : laya→rule-judge\n 生成后端 : scripted（脚本化，设 DEEPSEEK_API_KEY 可换真实 LLM）\n 判定放行：list_dir（auto）\n 判定放行：read_file（auto）\n 判定 12 次 50.4ms（均 4.2ms）\n 模型 1 次 600.9ms\n 判定 : 模型 = 12.0 : 1 判定耗时只占 7.7%\n```\n**同一个命令在不同环境下自动走不同的判定后端，数字也完全不同。** 下面是三种环境的实测：\n| 你的环境 | `npm run demo` 实际的后端 | 判定 : 模型 | 判定耗时占比 |\n|---|---|---:|---:|\n| 全新 clone（无 key） | `laya→rule-judge` | 12 : 1 | **7.7 %** |\n| 有 `TYPESAFE_API_KEY` | `jev→laya→rule-judge` | 13 : 1 | **79.4 %** |\n| 本地 Laya sidecar 在跑 | `laya→rule-judge` | 12 : 1 | ~38 % |\n> 上面那组头条数字来自**规则表，不是模型** —— 它演示的是「loop 结构长什么样」，\n> 不是「判定有多准」。真实判定的质量与延迟见\n> [Which decision backend](#which-decision-backend-and-what-it-costs-you)。\nZero dependencies. Zero build step. Runs offline with no API key.\n---\n## The problem\nTake a task that needs two tool calls. A conventional agent burns a model call on each of these:\n| Question the loop asks | Conventional agent | JevLoop |\n|---|---|---|\n| Do I need to act yet? | LLM call | decision |\n| Which tool? | LLM call | decision |\n| Is this call safe? | LLM call, or nothing at all | decision |\n| Did it work? | LLM call | decision |\n| Am I done? | `max_iter` counter | decision |\n| Can I ship this answer? | **nothing** | decision |\nYou were paying generation prices for decisions. A decision is one forward pass over a fixed candidate set — no tokens generated, nothing to parse, ~10–40 ms on a GPU.\n## Quick start\nNeeds **Node ≥ 22.6** (it runs TypeScript directly, no build).\n```bash\ngit clone https://github.com/Xubqpanda/JevLoop\ncd JevLoop\nnpm run demo\n```\nThat's it. No `npm install`, no API key, no network — the demo falls back to a deterministic rule judge so the whole loop runs offline.\n**Use a real decision model:**" + }, + "elberacasa/omawish": { + "sha": "bb2be38501638237b1b16ea70f1302b48b78861d", + "path": "README.md", + "size": 17950, + "excerpt": "

\n \"Omawish:\n

\n

A System One for Omarchy.

\n

Type what you want, in your own words. Omarchy does it.
\nOn your machine: no account, no key, no network, 35 milliseconds.
\nTrained in the open on one RTX 3080 Ti.

\n

\n \"Tests\"\n \"Trained\n \"60\n \"Zero\n \"35\n \"Licence\"\n

\n---\nOmarchy can do about four hundred things, each behind a command, a menu or a keybinding you have\nto know. Omawish is one line of text in front of all of them, and in front of your apps, your\nopen windows and anything you teach it. You type *make my screen warmer*, *turn bluetooth off*,\n*open cursor*, *remind me in 20 minutes to call mom* or *take a fullscreen screenshot and copy\nit*, and the right thing happens.\nIt is not a chatbot, and it never writes a command. A **System One** model does one small thing:\nshown what your desktop can do, it says how likely each is to be what you meant. Plain, tested\ncode decides what that is enough for.\n| The model says | Omawish does |\n|---|---|\n| one thing, clearly, and it is harmless | runs it, and gets out of the way |\n| one thing, clearly, but it changes your system | names it and waits for Enter |\n| it could be a few | offers them with their probabilities |\n| it is clear what, but not with which value | asks for the rest |\n| it is a question an action can answer | shows the answer in the bar; Enter copies it |\n| nothing on your desktop does that | says so, and Enter searches the web for it |\n

\"Typing

\n

It reads the wish while you type it. The model answers in 30 milliseconds, so what Enter will do is on screen before you press it.

\n

\"Typing

\n

It knows how things stand. You see what Enter will change, or already 4000 K when there is nothing to do. Every switch goes both ways: less warm, bring the bar back, turn notifications back on.

\n\n\n\n\n\n\n\n\n" + }, + "early-effect/hexis": { + "sha": "4ffae17a7060ebc510508b4b5aab48b475fdcc96", + "path": "README.md", + "size": 1468, + "excerpt": "# Hexis\nZIO / Scala 3 SDK for [TypeSafe System One](https://docs.typesafe.ai/introduction) (Jev).\nSend a state and typed questions. Get structured answers, probabilities, and confidence.\nCompose those answers in your code.\nCross-built for JVM, Scala.js, and Scala Native. HTTP is [heddle](https://github.com/early-effect/heddle) 0.4.0+.\nDocs: [earlyeffect.rocks/hexis](https://www.earlyeffect.rocks/hexis/) (after the first Pages deploy).\n## Install\n```scala\nlibraryDependencies += \"rocks.earlyeffect\" %% \"hexis\" % \"0.0.0\"\n// Scala.js / Native\nlibraryDependencies += \"rocks.earlyeffect\" %%% \"hexis\" % \"0.0.0\"\n```\nSet `JEV_API_KEY` (or `TYPESAFE_API_KEY`).\n## Quick start\n```scala\nimport hexis.*\nimport zio.*\nenum Team derives ChoiceDomain:\n case Billing, Technical, Sales\nval questions = (\n department = Choice[Team](\n \"Which team should handle this?\",\n Team.Billing -> \"Payment or subscription issues\",\n Team.Technical -> \"Bugs or integration problems\",\n Team.Sales -> \"Pricing or account questions\",\n ),\n urgency = Noul(\"The message conveys urgency or time-sensitivity\"),\n)\nval run =\n for\n cfg <- ZIO.fromEither(Config.fromEnv())\n out <- SystemOne.evaluate(\"Help, payouts have been failing for 3 days.\", questions)\n yield (out.department.choice, out.urgency.noul)\n// provide Transport.live ++ ZLayer.succeed(cfg) ++ SystemOne.layer\n```\nTests use `Transport.test(script)` and never hit the network.\n## License\nApache-2.0" + }, + "caohy1988/jev-guard-smoke": { + "error": "gh: Not Found (HTTP 404)" + }, + "dinkarjuyal/jev-gepa": { + "sha": "655920a0144af6584d044638a77b83f3c81c1cde", + "path": "README.md", + "size": 6297, + "excerpt": "# Jev + GEPA: fast NLI judges as an in-the-loop training signal\nWiring [Jev](https://huggingface.co/AlexWortega/openjev) (a small, fast, local NLI cross-encoder) into [GEPA](https://arxiv.org/abs/2507.19457) (a real, published reflective prompt optimizer, ICLR 2026 Oral), to test whether cheap per-step diagnostic tags improve GEPA's reflection step over raw trace text alone.\n**The idea.** LLM judges are usually too slow/expensive to run inside a training or optimization loop — they show up at evaluation time, after the fact. Jev is small and local enough (~9GB, one GPU, no per-call API cost) to plausibly tag every step of every rollout. This repo tests whether that lets a reflective prompt optimizer's feedback channel carry a real, structured, cheap semantic signal instead of just raw text.\n## Layout\n```\nadapters/ GEPAAdapter implementations (baseline vs. Jev-enriched, per task)\ndrivers/ scripts that actually run gepa.optimize() for each experiment\nresults/ real result JSONs and captured example traces from completed runs\ndocs/ supporting data (e.g. the offline chunk-context validation set)\n```\n- `adapters/aime_gepa_adapter.py` — AIME (GEPA's own paper benchmark). Contains the two most important pieces of engineering in this project: `_chunk()` (splits a single chain-of-thought response into pseudo-steps for per-chunk Jev scoring, with a validated fix for representative sampling across a full trace and context-augmented premises — see below) and `_with_hard_timeout()` (an externally-enforced call timeout, added after litellm's own `timeout=` parameter failed to fire on two real, separate hangs).\n- `adapters/alfworld_gepa_adapter.py` — ALFWorld (embodied agent benchmark; the Jev-enriched arm here was lost to a real GPU hardware fault, documented as an honest negative result).\n- `adapters/debug_gepa_adapter.py` + `debug_tasks.py` — the original small pilot task (6 hand-verified buggy-Python-function tasks).\n- `drivers/run_baseline_only.py` / `run_jev_only.py` — the current, correct pattern: run each arm of the A/B comparison as an **independent, concurrent process** (not sequential steps in one script), since the two arms are fully independent and running them concurrently roughly halves wall-clock time for a paired comparison.\n## What was found, in order\n1. **Debug-task pilot**: a real positive result at tiny scale (n=2 val) — a pilot, not a result on its own.\n2. **ALFWorld**: baseline arm valid, Jev-enriched arm lost to a real CUDA ECC hardware fault. No comparison obtained.\n3. **AIME, small scale**: three consecutive runs saturated at 0.0 everywhere, traced to two real bugs — a reasoning model's `content` field silently truncated by a `max_tokens` cap before it ever reached its formatted answer, and a dataset-answer-format mismatch (`\"### 073\"` vs a correct-but-unpadded `\"### 73\"`) defeating naive substring grading. Both fixed.\n4. **AIME, full scale (v1)**: baseline appeared to win. Investigating why found a *third* bug — `_chunk()` truncated every response to its first 12 sentences, so tags describing late-trace behavior (verifying work, committing to a final answer) were structurally almost never sampled once real responses ran 15-38K+ characters.\n5. **AIME, full scale, chunking fixed**: Jev-enriched won 3x (0.6 vs 0.2) at n=15 validation.\n6. **AIME, replication check (n=30)**: the 3x margin did *not* replicate — near-parity, baseline nominally ahead (0.4 vs 0.367) — a real reversal at double the sample size, run specifically to stress-test the n=15 result rather than trust it. Per-problem analysis showed the two arms are still genuinely different (not interchangeable): of 5 disagreements on 30 problems, Jev-enriched was right and baseline wasn't on 4, the reverse on 1.\n7. **Offline diagnosis + context-augmentation fix**: analyzing the real captured Jev tags from run 6 (for free, no GPU) found 6 of 8 diagnostic tags were near-dead weight (e.g. `made_concrete_progress` never exceeded 0.29 confidence across 720 real chunks). The pattern: the one strong tag (`made_arithmetic_step`) is a literal, self-contained claim; the weak ones are pragmatic/functional claims that need to know what came *before* a chunk to judge — but each chunk was being scored with zero context. Validated on the same 720 real chunks (bare vs. chunk-plus-preceding-context as premise): nearly doubled the overall confident-tag rate (5.5%→8.3%) and helped every pragmatic tag substantially.\n8. **Infra hardening**: two real, separate hangs (73+ minutes and 15+ minutes, both with zero CPU activity) revealed that litellm's own `timeout=` parameter was not reliably firing. Replaced with `_with_hard_timeout()`, an externally-enforced timeout via a worker thread the caller can walk away from regardless of what the underlying call does — verified in isolation before redeploying.\n## Reusable engineering lessons\n- litellm strips everything before the first `/` in a model ID as a routing prefix — a self-referential ID like `openai/gpt-oss-120b` needs doubling (`openai/openai/gpt-oss-120b`) to survive.\n- GEPA's `reflection_lm` needs a plain callable, not a bare model string, to route through a custom `api_base`.\n- A reasoning model's `content` field can come back empty under a `max_tokens` cap that looks generous, because the model spends its budget on a separate `reasoning_content` field first.\n- Dataset answer strings should never be trusted to match a model's literal output format without checking.\n- Chunking a long generated trace for per-step scoring must sample representatively across the whole trace, not truncate to a prefix.\n- Pragmatic/functional NLI claims (\"commits to\", \"verifies\", \"abandons\") need surrounding context to score meaningfully; purely literal, content-checkable claims don't.\n- litellm's own `timeout=` is not something to trust blindly for hang protection — wrap remote calls in an externally-enforced timeout if a hang would be costly.\n- Run independent comparison arms as concurrent processes, not sequential steps in one script.\n## Status\nFull experimental writeup (papers with all figures, tables, and example traces) lives in Google Docs; ask the repo owner for the current link. This repo has the actual code and result data behind that writeup." + }, + "luckberonne/mini-jev": { + "sha": "795ab56ea245d37dde4c362999a6cc9a42c7d602", + "path": "README.md", + "size": 2630, + "excerpt": "# mini-jev\nClasificador de comandos de shell de \"una sola pasada\", inspirado en Jev (TypeSafe AI): no genera\ntexto, devuelve probabilidades sobre etiquetas fijas. Decide si un comando es:\n- `solo_lectura`: no cambia nada (`ls`, `git status`, `docker ps`…)\n- `reversible`: cambia algo que se puede deshacer (`mkdir`, `git commit`, `docker restart`…)\n- `destructivo`: puede perder datos o dejar el sistema inutilizable (`rm -rf`, `dd`, `DROP TABLE`…)\nSirve de guardián para cualquier programa que ejecute comandos (paneles, agentes, scripts).\nNo depende de ninguna máquina ni programa concretos.\n## Instalación y entrenamiento\n```bash\npython -m venv .venv && source .venv/bin/activate\npip install -r requirements.txt\npython src/gen_dataset.py # genera data/train.jsonl\npython src/train.py # entrena y guarda ./model (GPU si hay, si no CPU)\npython src/evaluate.py # mide sobre data/test*.jsonl\n```\nCon una GPU pequeña o sin GPU funciona igual: el modelo base tiene ~22 M de parámetros.\n## Uso\n```bash\npython src/cli.py check \"rm -rf ~/Downloads\" # advertir: parece destructivo\npython src/cli.py check \"ls -la\" --json # detalle con probabilidades\nuvicorn src.serve:app --port 8000 # API: POST /decide, POST /verdict\n```\nCódigos de salida de la CLI: `0` permitir · `10` confirmar · `20` advertir. Así se integra en\ncualquier flujo: `jev check \"$cmd\" && eval \"$cmd\"`.\n`verdict()` solo devuelve `permitir` si el modelo dice lectura, con confianza ≥ 0.9 y todas las\nherramientas del comando están en `src/known_tools.py`. Un binario desconocido siempre pide\nconfirmación, porque el clasificador solo generaliza con lo que vio al entrenar.\n## Configuración\n| Variable | Efecto |\n|---|---|\n| `JEV_MODEL` | carpeta del modelo entrenado (por defecto `./model`) |\n| `JEV_BASE` | modelo base de Hugging Face para entrenar |\n| `JEV_EXTRA_TOOLS` | archivo con tus propias herramientas (una por línea) que cuentan como conocidas |\n`--threshold` (CLI) o `?threshold=` (API) ajusta la confianza mínima.\n## Datos\n`gen_dataset.py` combina plantillas con rutas, nombres y paquetes variados y etiqueta por reglas\n(encadenar con `&&`/`;` toma la peor severidad; `| sh` o `| xargs rm` es destructivo).\n`data/test*.jsonl` son comandos escritos a mano, fuera del entrenamiento.\nPara adaptarlo a tu entorno, añade plantillas en `gen_dataset.py` y tus herramientas a `JEV_EXTRA_TOOLS`.\n## Límites\nEs una ayuda, no una garantía: no entiende el contexto (`python script.py` puede hacer cualquier\ncosa) y solo cubre las herramientas y patrones con los que se entrenó." + }, + "sarathi-aiml/jevsql": { + "sha": "9c05b06bf11aac2ed94e953eb71fcfd415e56c5c", + "path": "README.md", + "size": 10324, + "excerpt": "# jevsql — text-to-SQL where the model never writes SQL\nA weekend experiment with [Jev](https://typesafe.ai/), TypeSafe AI's \"System One\"\nmodel that returns **typed, probability-calibrated decisions instead of text**.\n## How Jev differs from a normal LLM\nA generative LLM produces tokens: you send a prompt, it writes an answer one\ntoken at a time, and if you need structure you parse it out and hope. Jev\nnever generates text at all. You send your program's **state** plus a battery\nof **typed questions**, and it answers all of them in one parallel pass:\n```python\nr = client.system_one(\n state=\"I was charged twice for order A-104. Please refund the duplicate.\",\n questions={\n \"department\": Choice(instructions=\"Which team should handle this\",\n criteria={\"billing\": \"Payments/refunds\",\n \"technical\": \"Bugs\",\n \"sales\": \"Pricing questions\"}),\n \"frustration\": Score(instructions=\"How frustrated is the customer\",\n criteria=[\"Calm\", \"Frustrated but civil\", \"Angry\"]),\n \"wants_refund\": Noul(instructions=\"Customer explicitly asks for a refund\"),\n },\n)\nr.answers[\"department\"].choice # \"billing\"\nr.answers[\"department\"].confidence # 0.97 <- calibrated: right ~97% of the time\nr.answers[\"frustration\"].score # 1.2 (between \"Calm\" and \"Frustrated\")\nr.answers[\"wants_refund\"].noul # 0.94 (probability of yes)\n```\nThree primitives, that's the whole API: **Choice** (pick one of up to 255\noptions), **Score** (a position on an ordered scale), **Noul** (yes/no as a\nprobability). Your code branches on the answers the way it branches on any\nother value. The distinction in practice:\n| | Generative LLM (GPT-5, Claude, ...) | Jev (System One) |\n|---|---|---|\n| Output | Free text / JSON you parse | Typed values, no parsing |\n| Structure errors | 0.6–45% depending on model/mode | 0% by construction |\n| Confidence | Vibes (\"I'm fairly sure...\") | Calibrated probability per answer (RLCD-trained: 0.8 means right ~80% of the time) |\n| 10 questions | ~10x the output cost/latency | ~same latency as 1 (parallel pass) |\n| Latency | 2–6 s typical | 70–500 ms |\n| Price | $1.25–$10 /MTok in + output billed | $0.042 /MTok in, **output free** |\n| Good at | Composing anything: prose, code, SQL | Deciding: route, classify, rank, gate |\n| Can't do | Be cheap/fast/calibrated enough to run on every record | Generate anything — no text, no code |\nNeither replaces the other. Jev is the cheap, fast, calibrated decision layer;\ngenerative models remain the composition layer. The natural architecture is a\ncascade: Jev decides *what* to do on every request, a generative model is\ninvoked only for the fraction that needs one.\n## The idea I tried: text-to-SQL" + }, + "willgriffin/pi-fusion-matrix": { + "sha": "dfb8b07ad2c06d212db224e664601ec9bad42958", + "path": "README.md", + "size": 49676, + "excerpt": "# pi-fusion-matrix\nSeveral models deliberate; one answer comes back — and every step of how that happened, including\n**who got routed to what**, is configuration you can read and a run record you can audit.\nA [pi](https://pi.dev) extension — and one for [omp](https://github.com/can1357/oh-my-pi), the fork of the same\nstack — for multi-model deliberation. One codebase, both harnesses. It owns resolution, routing, and\nexecution: which model chain answers each seat, whether a cheap decision can answer instead of a model\ncall, which shape a run takes, which fusion a request should even use in the first place, and whether a\ncoding turn deliberates at all or goes to one model. It\nregisters one model per fusion (`fusion-matrix/best`, `fusion-matrix/cheap`, …), runs each seat\nthrough the harness's own provider runtime and credential store, and returns a normal\nassistant-message stream. Providers, endpoints, credentials, transport, and accounting stay the\nharness's.\nStatus: implemented and verified on both harnesses — eight fusions register and run on pi\n0.84.2/0.85.1 and on omp 18.2.6, from one codebase, and a pinned rung answers agent turns with the\nharness's own tools. The twenty-five verification items in\n[`docs/plan.md`](docs/plan.md) carry their evidence inline, and the repository's own offline contracts\n(`scripts/interp-check.mjs`, `scripts/doctor.mjs`, both probes) pass with no keys and no network.\n## Why this exists\nTwo failures of the obvious implementation:\n- **Hardcoding is the default failure mode.** A fusion that pins one provider, one key, and five bare\n model ids cannot express \"the same model on a second billing route\", and every vendor rename becomes\n a code change. Here, providers are pi's own ids and the model chain is `matrix.json`, so a vendor\n release edits one field and no code.\n- **Silent degradation is the dangerous one.** A pipeline that quietly substitutes a cheaper model, or\n streams two models' answers into one message, or presents progress lines as a result, is worse than\n one that fails. Every substitution, skipped stage, failed seat, truncated decision state, and\n declined route here is reported in the stream and recorded in the run's `details`.\n## Install\n```bash\npi install git:github.com/willgriffin/pi-fusion-matrix\n```\nOr point pi at a checkout — useful while editing config, since a local path is not copied:\n```bash\npi install /absolute/path/to/pi-fusion-matrix\npi -e /absolute/path/to/pi-fusion-matrix # try it for one run, installs nothing\n```\nThen `/reload` in a running session. A git install is pinned to the ref you installed — pin one\nexplicitly (`…@v1`, `…@`) if you want a fixed point; `pi update --extensions` reconciles the clone\nto that ref rather than moving it. It declares no runtime dependencies — pi bundles the peers it lists\n— so there is nothing for the installer to fetch.\nThe same directory loads in **omp**, which reads the same pi-style extension API:\n```bash\nln -s \"$PWD/extensions/pi-fusion-matrix\" ~/.omp/agent/extensions/pi-fusion-matrix # omp's own extension dir\nomp --extension \"$PWD/extensions/pi-fusion-matrix/index.js\" -p \"…\" --model fusion-matrix/cheap # or for one run\n```" + }, + "sable-inc/jev-linter-action": { + "sha": "b1ac97554b8269cacca8c0989cf402aa235e593d", + "path": "README.md", + "size": 4791, + "excerpt": "# Jev Linter Action\nReview repository files using your own yes/no questions. [TypeSafe Jev](https://docs.typesafe.ai/api) returns probabilities; the action passes only when every expected answer meets its configured threshold. Runs on Node 24 with a checked-in bundle; consumers need no installation step.\n```yaml\npermissions:\n contents: read\nsteps:\n - uses: actions/checkout@v4\n - uses: sable-inc/jev-linter-action@v1\n with:\n model: jev-1.13.0\n glob: 'prompts/**/*.md'\n questions: |\n - Are these instructions internally consistent after honoring explicit overrides?\n - Do these instructions avoid duplicating the same behavioral rule?\n - Are navigation paths grounded in supplied evidence rather than guessed?\n api-key: ${{ secrets.TYPESAFE_API_KEY }}\n```\nNo configuration file is needed. `questions` is a YAML list inside a workflow\nblock scalar (`|`). A plain question expects **yes**, with probability at least\n**0.8**. Write questions as properties that should hold. For negative questions\nor custom thresholds, use an object:\n```yaml\nquestions: |\n - id: contradictions\n question: Do these instructions contain contradictory requirements?\n expect: false\n minProbability: 0.85\n - Does the prompt respect the user's requested scope?\n```\n`glob` accepts one pattern or a newline-separated list. All matched files are\nreviewed together by default; `per-file: true` reviews each file independently.\nMissing inputs, malformed questions and ambiguous combinations fail before calls.\nFor multi-suite configurations, the existing `config: .jev-lint.json` input\nremains supported (and local CLI file arguments still work). Do not mix `config`\nwith inline inputs. Its JSON schema is:\n```json\n{\"model\":\"jev-1.13.0\",\"suites\":[{\"name\":\"Prompts\",\"files\":[\"prompts/*.md\"],\"questions\":[{\"id\":\"consistent\",\"question\":\"Are these instructions consistent?\",\"expect\":true,\"minProbability\":0.8}]}]}\n```\nEach suite sends all matching files together, preserving filenames, so it can find conflicts across files. Set `\"perFile\": true` to evaluate each matched file independently, useful when each file contains an assembled agent configuration. Suites and question IDs must be unique. All patterns must match a file. Paths are relative to the repository root; imports outside the root, including symlinks, are refused.\nFor an expected `false`, the passing probability is `1 - P(yes)`. A probability of 0.5 fails; uncertainty needs review. Thresholds default to 0.8 for inline questions and must exceed 0.5. Calibrate questions and thresholds using labeled acceptable and violating examples. Pin a model version for repeatability; `jev-latest` is also accepted. The model can be wrong, and static lint does not measure how an agent behaves in a call.\nSelected file contents and questions are sent to TypeSafe. Select only files appropriate for that service; do not target credentials. The key is used only with the fixed HTTPS TypeSafe endpoint. The action does not execute target files or follow redirects. Use `pull_request`, not privileged execution of untrusted PR code. Fork workflows do not receive repository secrets; skip this job explicitly for forks or run offline checks there. PRs that can edit this action's configuration can change the rubric and targets; retain review for those changes.\nMissing keys/files/answers, malformed configuration, API errors, and over-limit input fail closed. Rate limits and transient HTTP failures retry up to three attempts with bounded backoff; each attempt times out after 30 seconds. Provider error bodies and target contents are not logged. Each request is limited to 512 KiB, with at most 128 files per suite and 16 MiB total for a per-file suite. These are byte guards, **not token counts**: respect your model's documented context window and split large documents at meaningful boundaries. Content is never silently truncated.\nOutputs: `passed` and `report` (a JSON report path). GitHub gets a step summary and failure annotations. Reports contain filenames, questions, model IDs, raw yes probabilities, thresholds, and verdicts, not target contents. Upload the report explicitly if retention is needed.\nLocal use:\n```sh" + }, + "Krug2/JevLM-Open": { + "error": "gh: Not Found (HTTP 404)" + }, + "Kwwwww74/OpenJev": { + "sha": "8f5ec58f92ea2ad2dddf9ad9e03e3153e38dc4a7", + "path": "README.md", + "size": 9, + "excerpt": "# OpenJev" + }, + "clouatre-labs/decisions-judge-mcp": { + "sha": "cd771f8872c011937bd41efdbf275669cfc0e6e8", + "path": "README.md", + "size": 3383, + "excerpt": "# decisions-judge-mcp\nA small MCP server exposing a single **`judge`** tool: send natural language plus JSON application state, get back **typed judgments** — yes/no probability (`noul`), choice among options, or score on ordered levels — in one fast request, with model/usage metadata. Failures return a `{fallback: true, error}` envelope instead of blocking, so it is safe to compose into agent workflows.\nCurrently backed by the [TypeSafe System One](https://typesafe.ai) model (Jev) via [`@typesafe-ai/sdk`](https://www.npmjs.com/package/@typesafe-ai/sdk). The provider-neutral `judge` primitive maps naturally onto related decision APIs such as [OpenRouter alphadecisions](https://openrouter.ai/docs/api/api-reference/alphadecisions/submit-a-decisions-questions-and-answers-request).\nFlagship consumer: [`clouatre-labs/agentic-coder-skill`](https://github.com/clouatre-labs/agentic-coder-skill).\n## Tool: `judge`\nInput:\n- `state` — JSON object or string (application state)\n- `questions` — map of name → `{type: \"noul\"|\"choice\"|\"score\", instructions, criteria?}`\n - `noul`: criteria omitted; answer is a yes/no probability\n - `choice`: criteria is `{option: description|null}`\n - `score`: criteria is an ordered array of level descriptions\n- `timeout_ms` — optional, max 60000\nOutput: `{answers, model, usage}` on success, `{fallback: true, error}` on failure.\n## Install\nRequires Node >= 20 and `TYPESAFE_API_KEY` in the environment (`TYPESAFE_AI_TOKEN` is accepted as a legacy fallback).\n```sh\n# recommended: run on demand via npx (no global install)\nnpx -y decisions-judge-mcp\n# or pin a version for supply-chain reproducibility\nnpx -y decisions-judge-mcp@1.0.2\n# or install globally\nnpm i -g decisions-judge-mcp\n```\n> **Tip:** unpinned `npx -y decisions-judge-mcp` always runs the latest published\n> version. For deterministic, supply-chain-hardened setups, pin an exact version\n> or install globally and update deliberately.\n## MCP client configuration\nAll examples use `npx -y`, the standard pattern for Node-based MCP servers\n(see Context7 and Brave Search). Swap in `decisions-judge-mcp@` in the\n`args` if you prefer pinning.\n**pi** (`~/.config/pi/mcp.json` or equivalent):\n```json\n{\n \"mcpServers\": {\n \"decisions-judge\": {\n \"command\": \"npx\",\n \"args\": [\"-y\", \"decisions-judge-mcp\"],\n \"env\": { \"TYPESAFE_API_KEY\": \"...\" }\n }\n }\n}\n```\n**Claude Code** — one command:\n```sh\nclaude mcp add --scope user decisions-judge \\" + }, + "YuanKJing/Jev-as-Policy": { + "sha": "2b83b5db78bd8d1bf4acdeb90fe8ab8e04b7948f", + "path": "README.md", + "size": 225, + "excerpt": "# Jev-as-Policy\nThe highly anticipated open-source repository for **JEV as Policy** enables one-click setup of the simulation environment. Evaluations of Astra + JEV on benchmarks such as RoboTwin will also be released soon." + }, + "cobusgreyling/Jev": { + "sha": "9de601117711b440fc64776052511a2dbc099240", + "path": "README.md", + "size": 9262, + "excerpt": "# Jev Showcase — TypeSafe System One, not chat\n

\n \"Jev\n

\n

\n Unofficial operator-level companion for TypeSafe Jev
\n Choice · Score · Noul · parallel fan-out · confidence as a second axis
\n Jev 1.13 · released 15 September 2026\n

\n

\n What people miss ·\n Announcement ·\n Docs ·\n Models ·\n Article ·\n Showcase framework\n

\n---\nMost posts stop at **“fast structured output.”** This repo surfaces the **docs-only details** that change how you build: Jev is a **function call**, questions run **in parallel**, **output is free**, **confidence is not probability**, and **code owns the workflow**.\n| Under-known fact | Why it matters |\n|------------------|----------------|\n| **Not a chatbot** | No prose, code, or explanations. Pair with a generative model. |\n| **Three primitives** | Choice / Score / Noul. Question IDs are not sent to the model. |\n| **Fan-out** | Many questions, one state, one call. ~12× cheaper than serial. |\n| **$0.042 / MTok, output free** | Optimize question design, not completion length. |\n| **Noul has no `confidence`** | Do not copy a Noul threshold onto a Choice. |\n| **No fine-tune** | Shape answers via `state` + `instructions` + `criteria`. |\n| **Literal + no math** | Count, dates, and arithmetic stay in code. |\nFull write-up: **[docs/WHAT-PEOPLE-MISS.md](docs/WHAT-PEOPLE-MISS.md)** · interactive cards in the lab.\n**Keywords:** TypeSafe, Jev, System One, RLCD, calibrated decisions, structured output, confidence routing, agent guardrails\n---\n## The one-liner\n> **Frontier-intelligence function call** — unstructured state in, typed probabilistic decisions out. 70–500 ms. Cannot invent a label off your schema.\n| | |\n|--|--|\n| **Model id** | `jev-1.13.0` (`jev-latest`) |\n| **Class** | System One (not a chat LLM) |\n| **Endpoint** | `POST /v1/systemone` |\n| **Context** | 64k / request; 32k for state + longest question |\n| **Input** | Text / JSON. No image, audio, or video |\n| **Output** | Choice · Score · Noul (+ probabilities) |\n| **Pricing** | **$0.042 / MTok input · output free** |\n| **Latency** | 70–500 ms (vendor) |\n| **Training** | RLCD — Reinforcement Learning for Calibrated Decisions |\n| **Released** | 15 September 2026 |" + }, + "zszz3/Pi-Jev-Guide": { + "sha": "45aaa093f82bf3cd81b93451e7b06d9bbbfd06d1", + "path": "README.md", + "size": 14302, + "excerpt": "# Pi Jev Guard\n原版 Pi Coding Agent 插件：按时机配置规则，并自带风险检查、输出脱敏、重复失败和缺少验证提醒。\n0.2.1 支持 `/jevguard login` 配置 API key，验证后立即生效。\n支持交互添加规则：安装后运行 `/jevguard add`，依次选择 **检查时机 → 本地匹配或 Jev 判断 → 命中后的动作**。无需修改插件源码。\n适配并测试：`@earendil-works/pi-coding-agent 0.85.1`，Node.js 22.19+。这是独立的实验性插件，不是 TypeSafe 或 Pi 官方产品。当前提供 Pi 适配器；DSH 适配器尚未实现，判断与追踪模块可复用。\n## 可以做什么\n| 场景 | 默认 guard 模式 |\n| --- | --- |\n| 凭据上传，例如 curl 上传 `.ssh/id_rsa` | 不设置专门的本地拦截或确认；走普通 Jev／自定义规则，未配置时放行 |\n| 递归删除、丢弃 Git 修改、强制推送、非传输类凭据文件访问 | 确认这一次操作；没有交互界面时拦截 |\n| 其他 shell、write、edit、自定义工具调用 | Jev 同时判断破坏性、未经授权的数据外传、偏离任务和规则冲突 |\n| 疑似违反 AGENTS.md 或偏离要求 | 给 Agent 一条可见提醒，不自动回滚文件 |\n| 相同参数产生三次相同错误 | 提示检查原因、换方法或说明阻塞，每种错误只在第三次提醒 |\n| 观察到源码或配置写入，之后没有成功验证记录 | 收尾时提醒执行相关检查，或说明为什么不适用／未执行 |\n| 工具结果包含已知凭据样式 | 本地替换文本及元数据中的凭据；可再用 Jev 判断残留凭据并隐藏结果 |\n`read`、`grep`、`find`、`ls` 的执行前检查只运行本地凭据路径规则。它们的结果仍接受输出处理。一次 Jev 动作检查同时发送四个问题；默认额外检查一次非空文本输出。默认配置下，未配置 API key 时不发送网络请求，只有本地规则、脱敏与证据追踪生效。自定义语义规则会执行自己声明的 onError。\n## 安装与试用\n直接从 GitHub 安装：\n```sh\npi install https://github.com/zszz3/Pi-Jev-Guide\n```\n已打开的 Pi 会话运行 `/reload`，然后运行 `/jevguard login` 配置 key，用 `/jevguard add` 添加规则。\n在这个项目目录安装运行依赖，然后登记为本地 Pi 包：\n```sh\nnpm install --omit=dev\npi install /absolute/path/to/pi-jev-guard\n```\n### 配置 API key\n在 Pi 里输入一个命令：\n```text\n/jevguard login\n```\n在隐藏输入框粘贴 TypeSafe API key，按 Enter。插件用一条固定测试请求验证，成功后保存并立即启用 Jev，后续启动自动读取，不需要环境变量或重启。Esc 取消；验证或保存失败时保留原配置。\n- `/jevguard status`：查看是否启用以及 key 来源，不显示 key。\n- `/jevguard login`：也可用来更换 key。\n- `/jevguard logout`：删除保存的 key，回到本地检查；自定义语义规则仍按 `onError` 处理。\nKey 保存在 Pi 用户目录的 `jev-guard/auth.json`（通常为 `~/.pi/agent/jev-guard/auth.json`），是权限为 `0600` 的明文凭据文件，不在项目配置或会话记录中。输入框只显示星号；不要把 key 放在命令参数中。验证请求只包含固定测试文本，不包含项目内容；启用后，正常语义检查会将脱敏后的相关上下文发送到 TypeSafe。\n无交互界面的自动化仍可使用 `TYPESAFE_API_KEY`。保存的 key 优先于环境变量；退出登录后，如环境变量仍存在，会回退到该 key 并明确提示。模型、超时和阈值环境变量的变更仍需重启 Pi。\n只想临时加载、不登记全局包，可以在安装依赖后运行：\n```sh\npi -e /absolute/path/to/pi-jev-guard/src/index.ts\n```\n可以先让 Agent 执行 `git status`，查看正常放行；再在专用临时测试目录中提出清理操作，观察确认提示。不要为了演示让它接触真实私钥或重要文件。\n## 命令\n```text" + }, + "laidick/system-one-benchmark": { + "sha": "5ffe72671473b1a0462e0dc7f05710e040eedc46", + "path": "README.md", + "size": 10650, + "excerpt": "# System-One Model Benchmark: Open-Weight vs. TypeSafe Jev\nComprehensive empirical evaluation comparing open-weight Local System-One decision models against **TypeSafe Jev 1.13.0** on an unseen holdout benchmark across 174 realistic software engineering tasks.\n## Executive Summary\nSystem-One models act as fast, zero-to-low-token semantic routers and risk scorers in agentic architectures. They classify intent (`task_class`), predict required reasoning depth (`reasoning_need`), gauge execution blast radius (`execution_risk`), and evaluate prerequisite binary conditions (such as `repo_understanding_required` or `small_local_model_sufficient`).\nThis benchmark evaluated six models under an identical canonical 7-question contract:\n1. **TypeSafe Jev 1.13.0** (Specialized Cloud API)\n2. **Decider-0.8B** (`Mapika/decider-0.8b`, Qwen3.5-0.8B backbone, 752M params)\n3. **Decider-2B** (`Mapika/decider-2b`, Qwen3.5-2B backbone, 1.88B params)\n4. **Local Qwen System-One V2** (Qwen3.6-35B-A3B MoE logit scoring)\n5. **Laya Typed-Decisions** (`convaiinnovations/laya`, ModernBERT-large 421M)\n6. **Laya Base** (`convaiinnovations/laya`, ModernBERT-large 421M)\n### Key Findings\n- **Decider-0.8B is the standout open-weight model**: Achieving **85.1% task_class accuracy**, it is statistically on par with TypeSafe Jev (89.7%, $p = 0.1859$ via McNemar test). It matches Jev on **debugging recall (88.2%)** and **research recall (100.0%)**, while beating Jev on simple coding recall (85.7% vs 78.6%), requiring only **~1.8 GB RAM**.\n- **Decider-2B offers top calibration**: Demonstrates the lowest Expected Calibration Error (**ECE 0.0314**) and beats Jev on binary gate Brier scores (`web_research_required` Brier = 0.0756 vs 0.0901).\n- **Substantial Local MoE Replacement**: Both Decider models outperform the 35B Local Qwen V2 baseline (80.5%) while reducing memory requirements from ~21 GB down to 1.8–3.8 GB.\n- **Where TypeSafe Jev Remains Essential**: Operations task classification (100.0% vs 76.7%–80.0%), scalar score calibration (MAE 0.47 vs 0.75 on reasoning need), and sub-300ms multi-question cloud fan-out latency.\n---\n## 1. Six-System Empirical Comparison Matrix\nTested on a frozen holdout benchmark of 174 software engineering tasks (`unseen_holdout_v1.json`, SHA256: `d8f5e8a9c0eae5885307e54beae403629b21d389adea54dfe813d10779bb4d44`).\n| Metric / Dimension | TypeSafe Jev 1.13.0 | Decider-0.8B | Decider-2B | Local Qwen V2 (35B) | Laya Typed (421M) | Laya Base (421M) |\n| :--- | :---: | :---: | :---: | :---: | :---: | :---: |\n| **task_class Accuracy** | **89.7%** (156/174) | **85.1%** (148/174) | 83.3% (145/174) | 80.5% (140/174) | 57.5% (100/174) | 52.3% (91/174) |\n| **Macro F1 Score** | **0.9005** | **0.8563** | 0.8367 | 0.8125 | 0.5037 | 0.4758 |\n| **Simple Coding Recall** | 78.6% (33/42) | **85.7%** (36/42) | **85.7%** (36/42) | 50.0% (21/42) | 47.6% (20/42) | 71.4% (30/42) |\n| **Debugging Recall** | **88.2%** (45/51) | **88.2%** (45/51) | 80.4% (41/51) | 78.4% (40/51) | 80.4% (41/51) | 37.2% (19/51) |\n| **Architecture Recall** | 90.3% (28/31) | 77.4% (24/31) | 87.1% (27/31) | **100.0%** (31/31) | 90.3% (28/31) | 87.1% (27/31) |\n| **Research Recall** | **100.0%** (20/20) | **100.0%** (20/20) | 85.0% (17/20) | 95.0% (19/20) | 15.0% (3/20) | 15.0% (3/20) |\n| **Operations Recall** | **100.0%** (30/30) | 76.7% (23/30) | 80.0% (24/30) | 96.7% (29/30) | 26.7% (8/30) | 40.0% (12/30) |\n| **Reasoning Need MAE** | **0.4724** | 0.7965 | 0.7461 | 0.5617 | 0.7605 | 0.9741 |\n| **Execution Risk MAE** | **0.6861** | 0.7855 | 0.7820 | 0.7874 | 0.8290 | 0.9115 |\n| **Ambiguity Level MAE** | 1.5682 | 0.8813 | 0.8543 | 1.5717 | **0.8506** | **0.7950** |\n| **repo_understanding Brier** | **0.1529** | 0.2371 | 0.1569 | 0.2096 | 0.2708 | 0.2920 |\n| **small_local_model Brier** | 0.1483 | 0.1499 | **0.1462** | 0.1968 | 0.2358 | 0.2526 |\n| **web_research Brier** | 0.0901 | 0.2329 | **0.0756** | 0.1243 | 0.1983 | 0.1826 |\n| **Calibration Error (ECE)** | 0.3297 | 0.0747 | **0.0314** | 0.3762 | 0.1008* | 0.0387* |\n| **$\\ge 0.90$ Conf Wrong Rate**| **1.4%** (2/142) | 6.4% (9/140) | **5.8%** (15/257) | 10.1% (16/159) | 0.0% (0/0)* | 0.0% (0/0)* |\n| **Repeatability Determinism**| **100.0%** | **100.0%** | **100.0%** | **100.0%** | **100.0%** | **100.0%** |\n| **1-Question Latency (Local)**| ~120 ms (API) | 239.3 ms | 214.9 ms | ~1,450 ms | 52.1 ms | **48.2 ms** |\n| **7-Question Fan-out Latency**| **~306 ms** (API)| ~2,589 ms | ~3,408 ms | ~1,562 ms | 208.0 ms | 186.1 ms |\n| **Max Native Context** | >4,096 tok | **32,768 tok** | **32,768 tok** | ~4,096 tok | 1,024 tok | 512 tok |\n| **Resident Memory** | 0 MB (Cloud) | **~1.8 GB** | ~3.8 GB | ~21.0 GB | ~1.6 GB | ~1.6 GB |\n| **Offline Autonomous** | No | **Yes** | **Yes** | **Yes** | **Yes** | **Yes** |\n*\\*Note on Laya Calibration:* Laya's zero high-confidence error rate reflects temperature flattening (`choice:3-5 = 1.76`) where no cases achieved $\\ge 0.90$ confidence on out-of-domain engineering tasks.\n---\n## 2. Statistical Significance (McNemar Tests)" + }, + "mallahyari/system-one-benchmark": { + "sha": "b4b8e76713c89ee429923a13c25577dd5c53f8cf", + "path": "README.md", + "size": 6836, + "excerpt": "# System One & Parallel Constrained Decoding Benchmark\n[![Buy Me A Coffee](https://img.shields.io/badge/Buy%20Me%20A%20Coffee-Support-FFDD00?style=flat&logo=buy-me-a-coffee&logoColor=black)](https://buymeacoffee.com/mehdiyari)\nEmpirical evaluation and benchmarking suite comparing:\n1. **TypeSafe Jev (`jev-1.13.0`)**: A frontier \"System One\" decision model via cloud API.\n2. **Local Open-Source PCD**: Parallel Constrained Decoding with `Qwen 2.5 1.5B (4-bit)` on Apple Silicon using Apple's **MLX** framework.\n3. **Local Autoregressive Baseline**: Standard token-by-token JSON generation with `Qwen 2.5 1.5B`.\nEvaluated against 50 real-world adversarial, jailbreak, and benign prompts from the LMSYS [`lmsys/toxic-chat`](https://huggingface.co/datasets/lmsys/toxic-chat) benchmark.\n---\n## Benchmark Results (50 Real-World Samples)\n| Metric | TypeSafe Jev (`jev-1.13.0`) | Local Open-Source PCD (MLX) | Local Autoregressive Baseline | What This Demonstrates |\n| :--- | :--- | :--- | :--- | :--- |\n| **Model Tier** | **Frontier System 1** | Open-Source 1.5B | Open-Source 1.5B | Frontier vs. Small Model |\n| **Execution Mode** | **Cloud API (HTTPS)** | Local Apple Silicon | Local Apple Silicon | Network vs. On-Device |\n| **Forward Passes** | **1 Pass (O(1))** | **1 Pass (O(1))** | ~30.8 Passes | **30x Compute Reduction** |\n| **Latency p50 (Median)** | **356.5 ms** (incl. network) | **227.2 ms** (on-device) | 735.3 ms (on-device) | Jev over HTTPS is 2x faster than local autoregressive |\n| **Latency Mean** | **353.6 ms** | **317.1 ms** | 836.6 ms | Parallelism slashes latency in half |\n| **Accuracy** | **84.0%** ⭐ | 52.0% | 54.0% | +32% accuracy jump from Jev's frontier RLCD training |\n| **Precision (Safety)** | **90.9%** (Only 1 FP!) ⭐ | 36.0% (16 FPs) | 38.5% (16 FPs) | Jev eliminates false alarms |\n| **Recall (Caught Attacks)**| **58.8%** | 52.9% | 58.8% | Caught identical attack vectors |\n| **F1 Score** | **0.714** ⭐ | 0.429 | 0.465 | 60%+ boost in balanced safety performance |\n| **Brier Score (Calibration)**| **0.1096** (Near-perfect) ⭐| 0.3884 (Uncalibrated) | N/A (Raw strings) | RLCD delivers true probabilistic honesty |\n| **Schema Syntax Errors** | **0.0% Guaranteed** | **0.0% Guaranteed** | 98.0% (1 Crash) | Mathematical 0% schema error |\n---\n## Key Findings\n1. **The Architectural Shift (Local PCD):**\n Evaluating structured decisions in a single parallel forward pass ($O(1)$) using KV-cache broadcasting and sub-vocabulary logit slicing reduces forward passes by **96.8%** (1 pass vs 31 passes) and speeds up on-device execution by **3.2x** (227ms vs 735ms) with **94.0% decision concordance** to full autoregressive generation.\n2. **The Training Shift (TypeSafe RLCD):**\n While open-source PCD proves the speed mechanism, TypeSafe's **Reinforcement Learning for Calibrated Decisions (RLCD)** delivers frontier quality: **84.0% accuracy**, **90.9% precision** (only 1 false positive out of 50 prompts), and a **0.1096 Brier score** showing true probabilistic calibration.\n3. **Catastrophic Schema Drift in Autoregressive Models:**\n On Sample #23, when presented with a prompt asking for a list of human behaviors, the autoregressive LLM suffered instruction confusion and hallucinated new JSON keys (`\"human_behavior\": [...]`), crashing downstream schema parsers. Both Jev and Parallel Constrained Decoding were mathematically immune to this error.\n---\n## Quick Start\n### 1. Installation\n```bash\ngit clone https://github.com/your-username/system-one-benchmark.git\ncd system-one-benchmark\npip install -r requirements.txt\n```\n### 2. Configuration (For Live TypeSafe Jev)\nCopy the example environment file and add your TypeSafe API key:\n```bash\ncp .env.example .env\n# Edit .env:\n# TYPESAFE_API_KEY=your_actual_key_here\n```" + }, + "ckaraca/awesome-jev": { + "sha": "3e58af61ccebd902207cc784eee84b8318d027d1", + "path": "README.md", + "size": 20131, + "excerpt": "# Awesome Jev [![Awesome](https://awesome.re/badge.svg)](https://awesome.re)\n> A curated list of tools, integrations, and experiments built on **Jev**, the System One model from [TypeSafe AI](https://typesafe.ai/) that makes fast, typed, confidence-aware decisions.\nJev doesn't write paragraphs. You give it a question and a set of candidates, and it returns a typed answer with a confidence score, in milliseconds and for fractions of a cent. That makes it a good fit for work that needs many small decisions in a loop: picking the next click in a browser or on a phone, routing a task to the right model, classifying documents, scoring code changes, or deciding whether to trade.\n**No waitlist needed:** Jev is available through [Vercel AI Gateway](https://vercel.com/ai-gateway) as `typesafe-ai/jev`.\nWithin each section, projects are sorted by GitHub stars. Counts are refreshed weekly by [a workflow](.github/workflows/stars.yml). Stars belong to the whole repository, including projects where Jev is one optional integration.\n## Contents\n- [Official resources](#official-resources)\n- [Frameworks & integrations](#frameworks--integrations)\n- [Browser & computer use](#browser--computer-use)\n- [Mobile & robotics](#mobile--robotics)\n- [Coding agents & developer tools](#coding-agents--developer-tools)\n- [MCP servers & agent skills](#mcp-servers--agent-skills)\n- [Observability](#observability)\n- [Data, search & classification](#data-search--classification)\n- [Trading](#trading)\n- [Games & fun](#games--fun)\n- [Apps with Jev inside](#apps-with-jev-inside)\n- [Community SDKs](#community-sdks)\n- [Open models & replications](#open-models--replications)\n- [Other lists](#other-lists)\n## Official resources\n- [TypeSafe AI](https://typesafe.ai/) - Product site for System One models and Jev.\n- [Documentation](https://docs.typesafe.ai/) - Guides, SDK references, patterns, and the HTTP API.\n- [Quick start](https://docs.typesafe.ai/introduction/quickstart) - From an API key to your first typed decision in Python or JavaScript.\n- [Primitives](https://docs.typesafe.ai/primitives) - Choice, Score, and Noul, and when to use each.\n- [Patterns](https://docs.typesafe.ai/patterns) - Confidence-gated routing, composite scoring, speculative fan-out, intent routing.\n- [TypeSafe Console](https://console.typesafe.ai/) - API keys and live request inspection.\n- [Workflow evals](https://evals.typesafe.ai/) - Published workflows, model comparisons, and methodology.\n- [skills](https://github.com/typesafe-ai/skills) - Official agent skills for designing TypeSafe workflows from Claude Code, Codex, and similar agents. ⭐ 272\n- [typesafe-sdk-js](https://github.com/typesafe-ai/typesafe-sdk-js) - Official TypeScript/JavaScript SDK with inferred answer types. ⭐ 129\n- [system-one-adapter-python](https://github.com/typesafe-ai/system-one-adapter-python) - Drop-in `TypeSafeClient` replacement backed by regular LLM APIs, handy for local testing. ⭐ 121\n- [typesafe-sdk-python](https://github.com/typesafe-ai/typesafe-sdk-python) - Official sync and async Python SDK. ⭐ 88\n## Frameworks & integrations\n- [ComposioHQ/composio](https://github.com/ComposioHQ/composio) - Python and TypeScript providers that let Jev select tools and bind supported arguments, with partial-call and abstention results. ⭐ 30.2k\n- [vercel/ai](https://github.com/vercel/ai) - AI SDK's TypeSafe provider exposes Jev through the experimental evaluation API for typed Choice, Score, and Boolean questions. ⭐ 26.8k\n- [0xPlaygrounds/rig](https://github.com/0xPlaygrounds/rig) - Rust agent framework with an experimental `rig-typesafeai` crate for typed Jev judgments. ⭐ 8.7k\n- [ax-llm/ax](https://github.com/ax-llm/ax) - DSPy-style framework with TypeSafe/Jev support for boolean and class signatures, plus a native client for probabilities and scoring. ⭐ 2.9k\n- [agentjido/req_llm](https://github.com/agentjido/req_llm) - Elixir library with typed Jev evaluation through TypeSafe or OpenRouter, retaining probability distributions and provider responses. ⭐ 577\n- [ash-project/ash_ai](https://github.com/ash-project/ash_ai) - Maps Ash action arguments and return types to Jev evaluation requests through ReqLLM. ⭐ 189\n- [donvito/ai-backends](https://github.com/donvito/ai-backends) - API server with a dedicated Jev evaluation endpoint and an interactive playground for typed decisions. ⭐ 145\n## Browser & computer use\n- [trycua/cua](https://github.com/trycua/cua/tree/main/libs/cua-driver/examples/jev-use) - Open-source computer-use platform with cross-OS drivers. Its `jev-use` example lets Jev choose the next action. ⭐ 23.4k\n- [browser-use/jev-ultrafast](https://github.com/browser-use/jev-ultrafast) - Browser agent that uses Jev to pick target elements and only calls a small LLM when it has to type text. ⭐ 5.6k\n- [awlevin/typesafe-computer-use](https://github.com/awlevin/typesafe-computer-use) - macOS computer use for about $0.0002 a step: OCR the screen, let Jev pick the next click. ⭐ 223\n- [socai-io/socai](https://github.com/socai-io/socai) - Browser and computer-use agent tuned for social media research and content extraction. ⭐ 190" + }, + "promptgtm-shared/clay-jev-people-ranker": { + "sha": "40402cc861cb84d1ace9110cabe24405505dd536", + "path": "README.md", + "size": 12210, + "excerpt": "# TypeSafe JEV Lead Scoring with Clay CLI\nAn Agent Skill and Python workflow for **Clay lead scoring**, **B2B prospect qualification**, and **AI-powered people search ranking** with **TypeSafe JEV**.\n## Quick answer\nThis project uses the official Clay CLI to retrieve people who match deterministic filters such as job title, company size, industry, seniority, and location. It then uses TypeSafe JEV to make the harder semantic decision: does each person actually match the intended role and qualification criteria?\nThe included JEV use case targets current operating Founders and Co-Founders. It filters false positives such as founding investors, founding employees, former founders, and people working in a Founder's Office before downstream enrichment.\nThe workflow is designed for GTM engineers, RevOps teams, sales operations, CROs, BDMs, SDRs, recruiters, and developers building programmable lead-qualification systems.\n## What is JEV?\nJEV is TypeSafe AI's first public System One model. Instead of generating long-form text, it returns typed decisions and probabilities that software can evaluate directly.\nThis workflow uses two TypeSafe primitives:\n- **Choice** classifies the candidate as an operating founder, investor or board member, founding employee, founder-support role, or unclear.\n- **Noul** estimates whether the candidate clearly and currently holds an operating Founder or Co-Founder role.\nCode then applies explicit thresholds to those outputs. This makes JEV useful as a fast decision layer for classification, routing, scoring, verification, filtering, and reranking—not as a replacement for generative writing or multi-step reasoning.\nOfficial background:\n- [Introducing System One Models and JEV](https://typesafe.ai/blog/introducing-system-one-models-and-jev)\n- [TypeSafe use-case map](https://docs.typesafe.ai/concepts/use-case-map)\n- [Choice primitive](https://docs.typesafe.ai/primitives/choice)\n- [Noul primitive](https://docs.typesafe.ai/primitives/noul)\n## JEV use case: Clay lead scoring and qualification\nClay is strong at hard filters. A Clay people search can narrow a market by title, geography, company size, industry, seniority, experience, and other supported fields.\nThe remaining problem is semantic ambiguity. A broad search for `founder` can also match titles such as:\n- Founding Investor\n- Founding Employee\n- Founder's Office\n- Executive Assistant to Founder\n- Former Founder\nThe workflow separates the jobs:\n```text\nClay Query Mode\n -> apply deterministic people and company filters\n -> return broad candidate pages\n -> JEV Choice + Noul qualification\n -> apply probability thresholds in Python\n -> export qualified people to JSON and CSV\n -> enrich only the selected rows\n```\nThis is a practical TypeSafe AI use case for GTM automation: Clay retrieves the market, JEV judges ambiguous fit, and Python controls the policy.\n## Measured founder-ranking test\nThe September 21, 2026 test targeted 500 US-based Founders or Co-Founders at Software Development companies using Clay's available company-size buckets.\n| Metric | Result |\n|---|---:|\n| Clay candidates evaluated | 577 |\n| JEV-qualified candidates | 472 |\n| Candidates rejected before enrichment | 105 |\n| Qualification yield | 81.8% |\n| Current-founder threshold | 0.90 |" + }, + "ashafizullah/jev-triage": { + "sha": "96dcb0a587985b7c90bac63b25a28b625c1b3643", + "path": "README.md", + "size": 16627, + "excerpt": "
\n# jev-triage\n**Automated issue & PR triage for open-source maintainers, powered by [Jev](https://docs.typesafe.ai) (TypeSafe AI).**\n[![CI](https://github.com/ashafizullah/jev-triage/actions/workflows/ci.yml/badge.svg)](https://github.com/ashafizullah/jev-triage/actions/workflows/ci.yml)\n[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](./LICENSE)\n[![Node](https://img.shields.io/badge/node-%3E%3D20-brightgreen.svg)](./package.json)\n[![TypeScript](https://img.shields.io/badge/TypeScript-strict-3178c6.svg)](./tsconfig.json)\n
\n---\nAn incoming issue or pull request is classified, scored, checked for missing information, compared against existing items for duplicates, and then labelled — with every uncertain decision routed to a human instead of guessed at.\n## What it does\n- **Classifies** the item: `bug`, `feature_request`, `question`, `documentation`, `spam`, `other`\n- **Scores impact** on a four-level rubric and derives a severity\n- **Checks completeness**: reproduction steps, version info, environment, expected behaviour\n- **Detects duplicates** against open items, using a local lexical prefilter plus Jev verification\n- **Applies labels and leaves one comment**, updating it on re-runs instead of posting again\n- **Notifies Slack, Discord or Telegram** for critical or spam findings\n- **Learns from corrections**: maintainer overrides are stored as ground truth for an accuracy report\n## Why Jev\nTriage decisions are a closed set — a category, a severity, a boolean — not free text. Jev answers typed questions against a state and returns typed answers with calibrated probabilities, in one parallel request, without generating a single token to parse. That makes it both cheaper and more reliable than coercing a large language model into emitting JSON.\n## How it works\n```\nGitHub webhook (issues / issue_comment / pull_request)\n │\n ▼ Probot app (signature verification, event routing)\n Ingestion ──► store a bounded audit record\n │\n ▼\n Preprocessor ──► clean the text, collect metadata (labels, author, first-time?, diff)\n │\n ▼\n Jev fan-out #1 ──► category · impact · completeness · spam signals · needs-human-review\n │\n ▼\n Duplicate candidates ──► one list call → lexical prefilter → top K\n │\n ▼\n Jev fan-out #2 ──► one Noul question per candidate → 0..1 duplicate score\n │\n ▼\n Decision engine ──► composite spam risk + confidence gates → planned actions\n │\n ├──► GitHub: labels, comment, close, assign\n └──► Slack / Discord\n │" + }, + "DolphinMiner/jev-rss": { + "sha": "caaf2f1ebb6a62052ee95d268c16fb568aa58838", + "path": "README.md", + "size": 8265, + "excerpt": "# Jev RSS\n[English](README.md) · [简体中文](README.zh-CN.md) · [Contributing](CONTRIBUTING.md)\n**A local-first RSS inbox with a semantic screening layer.**\nFollow your sources, describe what matters, and explicitly ask Jev to screen the noise. Keep every article, inspect the signals behind each judgment, and decide what to read yourself.\n![Jev RSS home page in English](docs/images/home-en.png)\n_Actual application UI with synthetic demo articles and preset judgments. This screenshot is not real news, a live Jev result, or an accuracy benchmark._\n## Why this exists\nReading everything is expensive in attention. Sending everything to a large-model summarizer can be expensive in API calls, too. Jev RSS explores a smaller step: ask bounded semantic questions first, then keep the evidence visible for human reading.\nThis is an early-stage **screening tool, not a fact checker or an automatic AI newsletter**. No measured accuracy, speed, or cost advantage is claimed.\n## What you can do\n- Subscribe to RSS / Atom feeds or an optional, independently hosted [RSSHub](https://github.com/DIYgod/RSSHub) instance.\n- Refresh feeds without calling a model. Screen explicitly, up to **20 articles in the current list** per batch.\n- Read **For you**, **Needs review**, and **Filtered**, with signals, rule versions, and judgment history.\n- Save articles, mark them read, and record feedback in a local SQLite database.\n- Optionally extract a public page's text, then screen again. Previous judgments remain in history.\n- Export curated RSS containing current-rule, real-model selections; demo and outdated judgments are excluded.\n- Use **English by default**, or switch to **简体中文** in the sidebar. The choice persists in this browser. Source content and your preferences are not translated.\n**Not included:** scheduled fetching, background screening, push notifications, automatic agent actions, LLM summaries, cross-source semantic deduplication, accounts, or a hardened public deployment.\n## Quick start\nRequires **Node.js 22.16+** and npm. Node 22 LTS is the reference runtime (`.nvmrc`). Docker is not needed unless you choose to run RSSHub.\nClone this repository and start the development server:\n```bash\ngit clone https://github.com/DolphinMiner/jev-rss.git\ncd jev-rss\nnpm ci\ncp .env.example .env\nnpm run dev\n```\nOn Windows, copy `.env.example` to `.env` with your file manager or `Copy-Item`.\nOpen [http://127.0.0.1:5197](http://127.0.0.1:5197). Vite proxies API requests to the local server on port **3847**. These ports must be available; stop your own previous instance if necessary.\n**No model key required to try it.** Add feeds, read, save, or choose **Explore the demo** to load explicitly synthetic samples. Missing credentials never produce fake judgments for real articles.\n### Connect Jev (optional)\nEdit the server-side `.env` and configure one provider:\n| Provider | Required | Model default |\n| ---------- | -------------------- | ---------------------------------------- |\n| Typesafe | `TYPESAFE_API_KEY` | `JEV_MODEL=jev-latest` |\n| OpenRouter | `OPENROUTER_API_KEY` | `JEV_OPENROUTER_MODEL=typesafe/jev-1.13` |\nRestart the API after changing environment variables. Typesafe takes precedence if both keys are set. OpenRouter is an alternative connection, **not** an automatic retry/failover path. Model availability and billing depend on your provider.\nKeep keys on the server. Never add a `VITE_` prefix, include them in screenshots, or commit `.env`.\n**Local-first does not mean offline inference.** Clicking Screen sends article text and reading preferences to your selected provider and may incur charges. Send only material you are allowed to disclose.\n### Your first reading session\n1. Add a source. Standard RSS / Atom URLs do not need RSSHub.\n2. Open **Reading preferences** and describe what to follow and skip.\n3. **Refresh feeds**, then explicitly **Screen** the current list.\n4. Read selections and spot-check Needs review / Filtered. Extract more text when an excerpt lacks context, then screen again." + }, + "Dreydrey9000/jev-relay": { + "sha": "2039013eb53d6098a6c2dd9ecb2c46b52c37462a", + "path": "README.md", + "size": 6375, + "excerpt": "# Jev Relay\n**Small decisions. Clear fallbacks. You keep control.**\nA local-first advisory router for Claude Code, Codex and Hermes/Jax. Routine choices use a configured local model. Important sanitized checks can use TypeSafe Jev. Unavailable providers and invalid answers return responsibility to the calling assistant, with no silent paid fallback.\nThis is an early experimental release. It routes **bounded decisions**, not entire coding jobs. Every answer requires review. No claim of Jev-equivalent local accuracy or automatic execution.\n![How Jev Relay works](docs/how-it-works.png)\n## Try it without a key or model\nPython 3.10+; the router itself has no third-party dependencies. The optional Laya-MLX environment requires Python 3.11+.\n```sh\ngit clone https://github.com/Dreydrey9000/jev-relay.git\ncd jev-relay\npython3 -m jev_relay.cli --input examples/route.json\npython3 -m unittest discover -s tests -v\n```\nWithout local configuration, the first command returns `status: review`, `provider: frontier`, reason `local_disabled`. **Frontier is a handoff to your calling assistant, not another API call.** A working fallback is the default, even without keys.\n## Install for your assistant\n```sh\npython3 scripts/install_skill.py --target codex\npython3 scripts/install_skill.py --target claude\npython3 scripts/install_skill.py --target hermes --hermes-home /path/to/hermes/profile\n```\nEnsure `~/.local/bin` is on PATH. Start a new assistant session if needed for skill discovery. Installation uses symlinks to this checkout; keep it at a stable path. Existing unrelated skills and launchers are preserved. The same CLI and skill contract is used across assistants. It does not replace their main models, private memory, quota routing, or permission systems.\n## Local Laya setup (Apple Silicon)\nSee the [pinned setup and evaluation commands](docs/local-setup.md).\nUse a dedicated environment with [Laya-MLX](https://github.com/mizorewww/laya-mlx). Our evaluation uses MLX 0.32.2 and `aac6fef/laya-mlx` revision `20aed815fc6acde75733882e7ec0e3f28aeb9717`. Install the optional runtime with `python -m pip install -r requirements-laya.lock` in a separate Python 3.12 environment. Download its checkpoint separately into a local directory. Keep model weights out of Git. The English model file is about 843 MB; runtime memory is larger. Published Laya context limits include questions and options.\nCreate `~/.config/jev-relay/config.json` using absolute paths:\n```json\n{\n \"local_enabled\": true,\n \"python\": \"/absolute/path/to/laya-venv/bin/python\",\n \"model_path\": \"/absolute/path/to/downloaded/laya-checkpoint\",\n \"guard_module\": \"/absolute/path/to/jev-relay/jev_relay/guard.py\"\n}\n```\nThe built-in guard supports macOS/Linux. The actual Laya-MLX worker needs Apple Silicon. A host can supply a stricter compatible guard; Drey's installation retains his existing guard. The worker is offline and never downloads weights. Long instructions, options, or state are rejected instead of silently truncated. Calls run with a process lock and hard timeout. On Mac they use background task policy and reduced priority. Kill switches, low disk, memory pressure and thermal warnings are respected. The built-in guard cannot inventory every third-party inference runtime; keep other models unloaded and use your host's stricter guard when available.\nTo disable local inference, create `~/.config/jev-relay/DISABLED`. Drey's existing `~/.codex/local-ai-router/DISABLED` is also honored. Never remove a watchdog kill switch without its owner's explicit approval. `--ack-resource-warning` acknowledges a currently reviewed soft warning only; it cannot bypass hard blocks.\n## Important hosted checks\nSet `TYPESAFE_API_KEY` through your normal secret manager/environment, or use macOS Keychain service `TYPESAFE_API_KEY` with your current user account. `JEV_API_KEY` is also supported for existing Hermes setups. Never commit or paste a key into a request.\n```sh\njev-relay --input examples/route.json --importance important --allow-cloud --sanitized\n```\nBoth flags are required. `--sanitized` is your assertion, **not automatic redaction**. The heuristic secret check is an extra tripwire and cannot guarantee privacy. Hosted requests go only to the fixed TypeSafe endpoint; redirects are rejected and billable calls are not automatically retried.\n## Result contract\n- `requires_review` is always true. The router never executes actions.\n- `status: review` means the calling assistant must evaluate evidence. Local outputs remain in shadow mode even when confident.\n- `status: advisory` is an important hosted check, still requiring review." + }, + "fruitymcdoo/JevChat": { + "sha": "abb03d83190aedc8055dda09aa1f79aa7dd2ef51", + "path": "README.md", + "size": 11345, + "excerpt": "# JevChat\nA chat interface on top of [Jev](https://docs.typesafe.ai), TypeSafe's decision model.\nJev never writes text, and there is no canned text either: every word of a reply is the result of\ntyped decisions over a 9,000-word dictionary, and a judge (also Jev) decides whether the reply is\ngood enough to send. The UI streams words as they are decided and shows every decision underneath.\n```\nuser> I just got a new puppy and she is adorable\n bot> Wow she's cute! What name is she? (accepted at 94%)\nuser> her name is Biscuit\n bot> Wow Biscuit! Adorable! (best effort, 87%)\n```\n## Run\n```\npython -m venv .venv\n.venv\\Scripts\\pip install -r requirements.txt\n# put TYPESAFE_API_KEY=... in .env\n.venv\\Scripts\\python app.py # http://localhost:5000\n.venv\\Scripts\\python smoke_test.py -v # scripted conversation with each word's decisions (--parallel for the other composer)\n```\nWithout an API key the app falls back to a keyword-overlap `MockDecider` so the wiring still runs.\n## How a reply is built\nThere are two composers, selectable in the UI. **Left to right** is the default and much the better one.\n### Left to right (`jevchat/linear.py`)\nEvery measurement below points the same way: Jev is very good at picking a whole word when it can\nsee a real sentence, and weak when the choice is removed from that. So the reply is written one word\nat a time, and every choice sees the real text so far.\n| Stage | What happens |\n| --- | --- |\n| plan | 1 request: what the reply should do, its tone, \"How many words should an ideal response to this query contain?\" (1-3 ... 30+), which sets the word limit, and whether the reply suits an emoji (yes for good news and casual chat, no for bereavement or facts). |\n| kind | What kind of word comes next: pronoun, helper verb, noun, verb, ..., a number, an emoji (if the plan said yes; at most 2, never adjacent, and free of the word limit), a word echoed from the user, punctuation, or stop. The top 3 kinds all go on. |\n| heats | A big kind is several flat lists of 255 words by frequency (nouns: 20 lists). Every list is asked at once and sends its best 3 words to a final. |\n| finals | One flat choice per kind among the heat winners. Its top 3 are nominated; with the thesaurus on, so are synonyms of its favourite. |\n| compare | The nominees rendered as whole texts (\"Sorry about your\", \"Sorry about that\", ...). Jev picks the one that reads best, or stops, or takes back the last word (undo, only at 60%+ probability). |\n| assess | 3 yes/no questions about the finished reply: `responds`, `grammatical`, `complete`. Accepted when `responds` reaches `ACCEPT_THRESHOLD` (default 0.90, adjustable per message in the UI) and `grammatical` is at least 0.5. |\n| rewind | On rejection the word Jev was least sure of is ruled out at its position and writing resumes from there, up to 3 attempts. Then the best attempt is sent, flagged as below the bar. |\nEvery composing prompt opens with the task context (who is speaking, what a reply is for), and every\nword question carries a hint not to reuse the user's words; see the parrot experiment below.\nTypical cost: about 4 waves and 20k input tokens per word. A short reply is 3-5 s; a two-sentence reply\nwith retries is 50-130 requests, 9-17 s and 250k-750k tokens (about $0.01-0.03).\n### Parallel slots (`jevchat/composer.py`)\nThe reply is a row of numbered **slots**, each holding one word, one punctuation mark, or nothing.\nEvery open slot is asked about at the same time, so a round costs the same few request-waves\nhowever long the reply is. (A wave is a set of independent questions over one state, split across\nconcurrent requests when it would not fit in one.)\n| Stage | What happens |" + }, + "javsanesq/jevlab": { + "sha": "87d810d6a4b2633840606f0c18dd2313feaaeff2", + "path": "README.md", + "size": 30431, + "excerpt": "# jev workbench/harness\nA place to try small AI judgments and understand their answers. Give it a customer\nmessage, for example, and ask which team should help, how disruptive the problem\nis, and whether the customer wants a refund. See the alternatives, uncertainty,\nand the rule for when a person should check the result.\nThe **Jev model** is made by TypeSafe. This independent **jev workbench** helps you\nuse it from your Mac's terminal. It saves reusable designs, results, and learning\nprogress. It does not send customer messages, issue refunds, or execute the\nmodel's suggested action. An optional coach can suggest clearer questions.\n**Start here:** [One-page quickstart](docs/QUICKSTART.md) ·\n[Complete beginner's guide](docs/GUIDE.md).\nThe guide assumes no terminal experience and includes installation, a free demo,\na real worked example, and troubleshooting.\n## Try it\nIf already installed, run one command at a time:\n```sh\njev tour\njev demo\njev guide\n```\nThe tour introduces the app. The demo is a clearly labeled illustrative recording\nstored on disk: no key, network request, or charge. The guide is available offline;\n`jev guide --web` opens a local browser copy. To use the live model, save your\nTypeSafe key in macOS Keychain through the tour or `jev config`.\nRun `jev` to open the workbench. **Ctrl+E** explains the focused control,\n**Ctrl+G** opens the glossary, **Esc** goes back, and **Ctrl+Q** quits.\nSimple mode is the default; **More options** reveals advanced controls.\nEvery form field has a label, explanation, example, and local validation where\nneeded. Field help is visible in Simple mode and expandable in Expert mode.\nSingle Jev runs start immediately. Results show latency, token usage, and cost\ncalculated from returned usage. Batch and eval prompt before starting, with a\nseparate **Don't ask again** choice for each. Other paid workflows, such as coach\nadvice and comparisons, still confirm. Estimates are not spending caps; JSON,\npiped, and unattended commands keep their existing scripting budget rules.\n## What you can do\n| Use | Start here |\n| --- | --- |\n| Try and edit a reusable decision design | `jev` |\n| Revisit saved results | `jev history` |\n| Practice with ten short lessons | `jev learn` |\n| Browse seven example patterns | `jev library` |\n| Test accuracy and tune human-review rules | `jev eval` |\n| Process a file of cases | `jev batch` |\n| Compare two designs | `jev compare` |\n| Ask an optional design coach | `jev coach` |" + }, + "andrest04/jev-lab": { + "sha": "741995e66b02d6a11aa07c40620189ced9d3a3e8", + "path": "README.md", + "size": 2063, + "excerpt": "# Jev Lab\nA local lab for learning and testing [TypeSafe](https://docs.typesafe.ai)'s Jev model.\nJev is a System One model: you send state and typed questions, and it returns typed\nanswers (probabilities), not prose. Your code stays in control of the workflow.\n## Run\nGet a key at https://console.typesafe.ai/keys, then either export it or put it in a\ngit-ignored `.env` file next to `server.mjs`:\n```bash\n# .env\nTYPESAFE_API_KEY=your-key-here\nPORT=4173 # optional\n```\n```bash\nnode server.mjs # http://localhost:4173\n```\nNo key? The lab starts in **demo mode**: a keyword heuristic that mimics the response\nshape so you can explore the interface. It is not Jev and its answers are not evidence\nof how Jev behaves.\n## What is in it\n- **Examples**: six worked patterns (ticket routing, composite scoring, function calling, semantic search, citation checks, guardrails). Each shows the typed questions, the answers as instruments, the decision your code makes, and the logic to copy.\n- **Playground**: write state, build noul / choice / score questions, run, and copy the request as curl, fetch or the Python SDK.\n- **Learn**: the mental model, primitives, an interactive confidence explorer, known pitfalls and limits.\nThree answer modes are always labeled in the UI: **live** (real Jev), **sample** (hand-written illustrations for preset inputs) and **demo** (keyword heuristic in the playground).\n## Languages\nThe UI is English and Spanish. Pick a language in the top bar; the choice is stored as `jev-lang` and also follows the browser on first visit.\nTechnical terms stay in English in both languages (`noul`, `choice`, `score`, `confidence`, Playground, pattern names, API/env names). Question instructions, criteria, preset state, and anything you type in the Playground stay in English: that is what Jev receives.\n## Test\n```bash\nnode --test\n```\n## Why a local server\nThe API key must stay server-side. The browser talks to `/api/systemone` on this\nserver, which forwards to `https://api.typesafe.ai/v1/systemone`." + }, + "koteitan/laya-bot-det": { + "error": "gh: Not Found (HTTP 404)" + }, + "kzkhykw/jev-or-not": { + "sha": "7cbe3a508fab5775105262379189e6f0011722d2", + "path": "README.md", + "size": 1120, + "excerpt": "# Jevる？\n入力データの性質から、利用するモデルや処理方法を選ぶための判断フローです。\n```mermaid\nflowchart TD\n A{\"賢さいらない？\"}\n B{\"リアルタイム\"}\n C{\"大量？\"}\n D{\"入力が非構造化
データ？\"}\n E{\"分類が動的？
教師データ用意
できない？\"}\n F{\"学習めんどい？\"}\n J[\"Jev\"]\n BERT[\"BERT\"]\n IF[\"if文\"]\n LIGHT[\"軽いLLM\"]\n SOME[\"ある程度のLLM\"]\n A -->|Yes| B\n A -->|No| SOME\n B -->|Yes| C\n B -->|No| LIGHT\n C -->|Yes| D\n C -->|No| LIGHT\n D -->|Yes| E\n D -->|No| IF\n E -->|Yes| F\n E -->|No| BERT\n F -->|Yes| J\n F -->|No| BERT\n```\n## 判断の流れ\n- 賢さが不要なら、ある程度のLLMを利用\n- リアルタイム性が必要なら、データ量を確認\n- 大量の非構造化データで分類が動的、または教師データを用意できない場合は学習の手間を確認\n- 学習が面倒ならJev、そうでなければBERT\n- 非構造化データでなければif文、リアルタイム性が不要なら軽いLLM" + }, + "peach-zhang/typesafe-go": { + "sha": "109ec34a82475c67df79d5ad2400d1314662d8e9", + "path": "README.md", + "size": 6520, + "excerpt": "# typesafe-go — TypeSafe System One 的 Go SDK\n参考官方 JavaScript SDK(`TypeSafeClient` + `choice()/score()/noul()` 辅助函数)设计的\nTypeSafe System One API(旗舰模型 Jev)Go 客户端。模型提供类型化的判断与概率,\n代码掌控工作流。\n## 安装\n```sh\ngo get github.com/peach-zhang/typesafe-go\n```\n```go\nimport typesafe \"github.com/peach-zhang/typesafe-go\"\n```\n## 项目结构\n```\ntypesafe-go/ # 包在仓库根,import 即模块路径\n├── client.go # Client、SystemOne、roundTrip、函数式选项\n├── questions.go # Noul / Choice / Score 问题构造\n├── response.go # SystemOneResult、Answer、Usage\n├── models.go # Models 端点与 ModelCard\n├── retry.go # RetryPolicy、指数退避、Retry-After\n├── errors.go # APIError / 连接 / 超时错误与判定函数\n├── logger.go # Logger 接口与 LogLevel\n├── typesafe_test.go # 单元测试(httptest 模拟服务端)\n└── examples/triage/ # 工单分类示例\n```\n## 快速开始\n```sh\n# Windows PowerShell\n$env:TYPESAFE_API_KEY = \"你的key\"\ngo run ./examples/triage\n```\n```go\nclient, err := typesafe.NewClient() // 读取环境变量 TYPESAFE_API_KEY\nresult, err := client.SystemOne(ctx, typesafe.SystemOneRequest{\n State: map[string]any{\"ticket\": map[string]any{\"message\": \"Help! Payouts failing for 3 days.\"}},\n Questions: typesafe.Questions{\n \"category\": typesafe.Choice(\"Which team should handle `ticket.message`?\", typesafe.ChoiceCriteria{\n \"billing\": \"Payments, invoicing, refunds\",\n \"technical\": \"Bugs, outages, integrations\",\n \"sales\": nil, // 无描述的选项\n }),\n \"is_urgent\": typesafe.Noul(\"Does `ticket.message` convey urgency?\", nil),\n \"frustration\": typesafe.Score(\"How frustrated is the customer?\",\n typesafe.ScoreCriteria{\"Calm\", \"Frustrated\", \"Very angry\"}),\n },\n})" + }, + "twilwa/pi-typesafe": { + "sha": "195563911db2f04799efe12a707a66e695eb46c3", + "path": "README.md", + "size": 20418, + "excerpt": "# pi-typesafe\nA Jev coding sidecar for Pi: four pre-flight hazard checks before `bash`, `write`,\nand `edit`, and four diff-quality checks after successful `write` and `edit`.\nQuestions share one TypeSafe System One request per phase. Jev returns typed\nvalues; this extension applies thresholds and writes fixed feedback text.\n**Advisory mode is the default.** It records assessments and appends qualifying\nfeedback after execution without blocking tools. Blocking requires explicit\nopt-in; explicit shadow mode records assessments without changing model-visible\nresults. All three modes use the same checks.\nMissing credentials or any judge failure always leaves the tool untouched.\n## Try it\nUse Node.js 24 and **Pi 0.85.1** (`@earendil-works/pi-coding-agent`). Pi itself\nrequires Node >=22.19.0; this repository uses Node 24 to run TypeScript tests\nwithout a build step.\n```sh\nnpm ci\nnpm run check\npi -e ./src/extension.ts\n```\nWith no key, the extension loads and does nothing. To enable evaluation, export\n`TYPESAFE_API_KEY` in Pi's environment, or copy `.env.example` to `.env`, enter the\nkey locally, and load it before starting Pi:\n```sh\nset -a\n. ./.env\nset +a\npi -e ./src/extension.ts\n```\nPi does not automatically load this repository's `.env`. Local `.env` files are\nignored by git; never commit credentials. The extension creates the existing\n`src/typesafe.ts` client only when a nonblank key is configured. Configuration is\nread when the extension loads; restart/reload it after changing configuration.\nFor a real-session trial, use a disposable Git repository, set `sidecar` to this\ncheckout's absolute path, and launch:\n```sh\nsidecar=/absolute/path/to/pi-typesafe\nmkdir jev-playground\ncd jev-playground\ngit init\npi -e \"$sidecar/src/extension.ts\"\n```\nAsk Pi to implement a small function and a test. Qualifying feedback appears in\ntool results, and assessments are recorded in the session's `jev-assessment`\ncustom entries. To record assessments without model-visible feedback, use:\n```sh" + }, + "alexwestco/llm-to-jev": { + "sha": "43cd94fba7527d78635b148628485c0cd1b1ad66", + "path": "README.md", + "size": 3637, + "excerpt": "# LLMtoJev\n**Turn decision-shaped LLM prompts into proposed Jev primitives.**\n[Try the live demo](https://alexwestco.github.io/llm-to-jev/) · [Read the Jev docs](https://docs.typesafe.ai/) · [Contribute](CONTRIBUTING.md)\nLLMtoJev finds bounded decisions inside prompts written for GPT, Claude, or Gemini and proposes Jev `Choice`, `Score`, and `Noul` questions. It is designed for classification, scoring, routing, and yes/no judgments that do not require generated prose.\nEach prompt is marked as fully convertible, partially convertible, or not convertible. For mixed prompts, the converter separates Jev-compatible decisions from writing, summarization, translation, and other generative work that must remain with an LLM.\n## Example\n**Input**\n```text\nClassify this support ticket as billing, technical, account access, or other.\nScore its urgency from 0 to 1, and decide whether it needs human review.\n```\n**Detected primitives**\n```json\n{\n \"suitability\": \"strong\",\n \"compatibility\": \"full\",\n \"questions\": [\n { \"id\": \"category\", \"type\": \"choice\" },\n { \"id\": \"urgency\", \"type\": \"score\" },\n { \"id\": \"needs_human_review\", \"type\": \"noul\" }\n ]\n}\n```\nThe interface exports ready-to-review examples for TypeSafe's official JavaScript and Python SDKs, plus the complete neutral JSON representation.\n## Why this exists\nMany LLM calls do not need prose generation. They ask a model to choose from known options, assign an ordered score, or answer a boolean question. Jev provides primitives for those bounded decisions. LLMtoJev helps identify prompts that may be candidates for that migration.\nThis is a conversion assistant, not an automatic guarantee of equivalent behavior.\n## Run locally\nRequires Node.js 20 or newer. There are no dependencies or build steps.\n```bash\ngit clone https://github.com/alexwestco/llm-to-jev.git\ncd llm-to-jev\nnpm run dev\n```\nOpen [http://localhost:3002](http://localhost:3002).\nRun the test suite:\n```bash\nnpm test\n```\nRun an optional live smoke test against Jev with your own API key:\n```bash\nTYPESAFE_API_KEY=your_key npm run test:live\n```\nThe key is read from the process environment and is never stored or printed.\n## Project structure" + } +} \ No newline at end of file diff --git a/research/archive/hourly/2026-09-21T01/revisit_high_this_run.json b/research/archive/hourly/2026-09-21T01/revisit_high_this_run.json new file mode 100644 index 0000000..8823058 --- /dev/null +++ b/research/archive/hourly/2026-09-21T01/revisit_high_this_run.json @@ -0,0 +1,24 @@ +[ + { + "id": "alexwestco/llm-to-jev", + "description": "Convert LLM prompts to Jev prompts", + "stars": 3, + "forks": 0, + "license": "MIT", + "language": "JavaScript", + "size": 29, + "pushed_at": "2026-09-20T12:06:01Z", + "created_at": "2026-09-19T15:20:36Z", + "updated_at": "2026-09-20T17:23:58Z", + "homepage": null, + "default_branch": "main", + "default_sha": "234058ab372d7754833c6279601755f8fda55d98", + "readme": { + "path": "README.md", + "sha": "43cd94fba7527d78635b148628485c0cd1b1ad66", + "size": 3637 + }, + "topics": [], + "desc_hash": "9f521cf7cc20" + } +] diff --git a/research/archive/hourly/2026-09-21T01/run_digest.json b/research/archive/hourly/2026-09-21T01/run_digest.json new file mode 100644 index 0000000..ca1d984 --- /dev/null +++ b/research/archive/hourly/2026-09-21T01/run_digest.json @@ -0,0 +1,11 @@ +{ + "hour": "2026-09-21T01", + "label": "1946", + "notes_section": "131", + "composition": "505-520", + "findings_batch": 113, + "revisit_high": 1, + "novel_high": 72, + "invented_signal": false, + "primary": "X-sentiment does not execute; 485 catalog \u2260 endorsement; AI-reviewed \u2260 gold; one-trial robot \u2260 Harbor; 10.59\u00d7 systems \u2260 ECE" +} diff --git a/research/changelog-hourly.md b/research/changelog-hourly.md index decb222..d01401f 100644 --- a/research/changelog-hourly.md +++ b/research/changelog-hourly.md @@ -17,6 +17,18 @@ This is the uniqueness-lock archive after hourly folds (#2–#40 / notes +## Hourly 1946 HIGH (notes.md §131 / items 505–520 / batch #113) + +- Fresh PR off latest `main` after merged #54 (aisearchio / §130) and merged #52 + (hourly 1843 / §129). Merged #54 owns 497–504 / #112. Skip Open-Jev #53. **HARD RULE:** do not reopen or amend + PR #23–#52. Does not bump 0.5.0. Skip Archer. Skip Open-Jev #53. + Skip #54 three. Quote *theirs*. No wrappers. `invented_signal: false`. +- X-sentiment does not execute. 485 catalog ≠ endorsement. + AI-reviewed labels ≠ gold. one-trial robot ≠ Harbor. + 10.59× systems ≠ ECE. Spanish −6.4 pp XNLI *theirs*. + llm-to-jev description rewrite. SHA unchanged. +- Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 + ## User-provided 1936 HIGH (notes.md §130 / items 497–504 / batch #112) - Fresh PR off latest `main` after merged #52 (hourly 1843 / §129) and diff --git a/research/notes.md b/research/notes.md index 55d8553..e6be55b 100644 --- a/research/notes.md +++ b/research/notes.md @@ -29495,6 +29495,15 @@ human review owns the question that ships. compiler as Jev. Do **not** treat exported SDK as a measured equivalent. Star counts ephemeral. + DENSIFY 1946 (description rewrite only; SHA unchanged). + GitHub description is still “Convert LLM prompts to Jev prompts.” + desc_hash `9f521cf7cc20` (was `653a47cc121d` in the 0940 lock + era). HEAD still `234058ab372d`. README SHA still `43cd94fb`. + **3★** (0940 uniqueness lock stays 2★; star-noise is not the + fold). desc rewrite ≠ SHA/behavior change. heuristic conversion + ≠ calibrated Noul. Do not mint a sibling first sighting. + + ### Adversarial review + testing hooks - Uniqueness-gate: the consecutive @@ -32873,3 +32882,319 @@ User-provided 1936 uniqueness lock: sgoedecke/system-one 20★ HEAD ebde2a2db706 **Open-Jev densify (`notes.md` §125).** DENSIFY the original 1441 card, not a sibling first sighting. HEAD 4933ee84951f README SHA ce1a587219e4. LoRA + scalar head + calibration temperature. not merged base models. customer-service P50 85.03 vs Jev 295.26 *theirs*. 1024/32 slower 1015.90 vs 301.37 *theirs*. systems latency ≠ semantic equivalence. Open-Jev TREC pending. hard acc ≠ calibrated Noul. type-valid ≠ exact. LoRA ≠ RLCD replica. Qwen/Qwen3.8-27B ≠ Archer. SHA move is not a replica. Do not reopen or amend PR #23–#52. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 +## 131. Hourly 1946 HIGH (2026-09-20 ~19:46 Boise / 2026-09-21T01:46Z) + +Measurement / class-member fold on a **fresh PR off latest `main`** +(`cursor/fold-hourly-1946-high-0532`) after `7a118b4` (merged #53 +Open-Jev densify on `notes.md` §125; merged #54 aisearchio census gaps, +`notes.md` §130 / items 497–504 / batch #112; merged #52 hourly 1843, +`notes.md` §129 / items 481–496 / batch #111; merged #51 hourly 1746, +`notes.md` §128). **HARD RULE:** do not reopen or amend PR #23–#54. +Do **not** re-fold §130 1936 / §129 1843 / §128 1746 *as a second census*. +Do not amend #53/#54. Do not mint a sibling Open-Jev first sighting +(merged #53 densified §125). This fold's IDs: `notes.md` §131 / +composition 505–520 / findings batch #113. + +Never reopen merged #7–**#52**. Densify `alexwestco/llm-to-jev` on +§118 only (description rewrite; SHA unchanged). Skip densify +`Zefan-Cai/Open-Jev` (merged #53 already densified §125). Skip `sgoedecke/system-one`, +`mithalouni/system-one-open`, `kotoba-lang/typed-decisions` (merged #54 / §130). +Skip already-carded `ikermoel/open-alternative-jev` (§49), +`nrdz-labs/fast-jev-opencode` (§62), `mallahyari/system-one-benchmark` (§61), +`liao96312/jev-arena-nanojev` (NanoJev namesake), `emirbartu/opencode-system-one`. +Quote READMEs. Mark *theirs*. No wrappers, keys, `npm` / +`pip` / `uv` / `docker` install recipes. `invented_signal: false`. +Hunches labeled. + +Lane is Augustus: **mathematical / logical / algorithmic mental models** +for Jev-class categorization/scoring across AI / SWE / **business / +knowledge work / life**, not SWE-only. PRIMARY this hour is **novel HIGH +measurement**: a crypto terminal that does not execute, independent +catalogs that are indexes, AI-reviewed comment labels that are not gold, +one seed-0 robot trial, a 10.59× systems timing that is not ECE, a +preregistered Spanish state audit. REVISIT densify is llm-to-jev +description rewrite only. Third-party benches stay *theirs*. Catalogs are +indexes. Soft scores ≠ hard gates. reconstruction ≠ replica. +Qwen3.8 ≠ Archer. Archer still **promised_not_landed**. + +Unique consecutive fragments (this hour) must appear as **one +substring** in overlays (see uniqueness gate): +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 + +### How-to-apply (five placements / measurement lenses) + +These are *class* lenses, not vendor tutorials. Same discipline as +§129 (reconstruction ≠ replica) and §118 (heuristic conversion ≠ +calibrated Noul). Formal methods **compose** with scoring: a Noul +is a SENSOR; a catalog is an index; a one-trial robot run is not Harbor; +AI-reviewed labels are not gold. + +1. **platform does not execute trades** + (*theirs*, brainstormity/Jev-X-Sentiment-Analysis). Quote *theirs*: + The platform does not execute trades automatically. It generates a + clear decision card with calculated entry ranges, stop losses, and + target levels so you can review the reasoning and execute manually. + Buy / Sell / Hold / Take Profit is a Choice on a dashboard, not a + fill. does not execute. Life analogue: a weather map is not a + thermostat. +2. **catalog ≠ endorsement / AI-reviewed labels ≠ gold** + (heyjunpenn/awesome-jev; NanmiCoder/jev-arena). Quote *theirs* + awesome-jev: an independent, community-maintained catalog of 485 + open-source projects. It is not affiliated with or endorsed by + TypeSafe AI. GitHub description said 503; the README badge is 485. + Quote the README. heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ + MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ + yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ + andyrewlee/awesome-system-one. Quote *theirs* jev-arena: 10,000 + comments; Jev 203.2s / $0.84 vs DeepSeek 823.5s / $1.50; three-way + 62.69% vs 67.26%; this is AI review, not a human gold standard; + strict 50.58% / 55.45%. AI-reviewed labels ≠ gold. *theirs* not Harbor. +3. **one-trial robot ≠ Harbor / 10.59× systems ≠ ECE** + (openroboto-ai/jev-robot-control; endman100/research-Qwen3.8-JevLike). + Quote *theirs* robot-control: one seed-0 trial per controller, not + success-rate estimates. Jev placed 113 cycles $0.018825 181.8s vs + Astra $5.93 707s vs mini 160-cycle limit. Quote *theirs* JevLike: + A JSON Schema 46.589s vs B binary 4.398s (10.59×); 71-way 3.465s; + 6 class flips; no labels; agreement ≠ accuracy; probabilities + uncalibrated; Qwen3.8 ≠ Archer. systems comparison ≠ semantic + equivalence. +4. **Spanish state costs accuracy and calibration *theirs*** + (marcosmartinez/jev-acento). Quote *theirs*: preregistered audit; + 19,200 calls, 3,200 paired, $0.58. XNLI 0.850→0.786 −6.4 pp. + ECE 0.057→0.101. p_max≥0.9 coverage 72.2% EN vs 63.4% ES. + A language swap on `state` is a measurement, not a replica claim. + *theirs* not Harbor. +5. **desc rewrite ≠ SHA/behavior change / does not execute** + (alexwestco/llm-to-jev densify §118; win4r/jev-skill-suggester; + PyModel/typesafe-mcp; Xubqpanda/JevLoop). Quote *theirs* llm-to-jev: + Convert LLM prompts to Jev prompts. SHA unchanged 234058ab372d. + heuristic conversion ≠ calibrated Noul. Quote *theirs* skill-suggester: + does not execute or install candidate skills. local_only ≠ Jev. + Quote *theirs* typesafe-mcp: evaluate is a stdio MCP server; the host + still reasons, edits, and executes. Quote *theirs* JevLoop: headline + 12:1 / 7.7% is from a rule table, not a model. rule-table ≠ model. + +### HIGH (revisit densify; keep original section ids) + +1. **[alexwestco/llm-to-jev](https://github.com/alexwestco/llm-to-jev)** + - DENSIFY §118 (MIT; **3★**; HEAD `234058ab372d`; README SHA + `43cd94fb`; desc_hash `9f521cf7cc20`). Description rewrite + Convert LLM prompts to Jev prompts. SHA unchanged. + desc rewrite ≠ SHA/behavior change. heuristic conversion ≠ + calibrated Noul. Do **not** mint a sibling first sighting. + Do **not** copy `npm`. + +### HIGH (novel) + +2. **[brainstormity/Jev-X-Sentiment-Analysis](https://github.com/brainstormity/Jev-X-Sentiment-Analysis)** + - NEW HIGH measurement (license null; **136★**; HEAD `5c932f941a92`; + README SHA `bf4134b44cda`). Quote *theirs*: platform does not + execute trades. Buy / Sell / Hold / Take Profit. Two-tier tweet + pipeline. does not execute. Do **not** copy keys. + +3. **[heyjunpenn/awesome-jev](https://github.com/heyjunpenn/awesome-jev)** + - NEW HIGH measurement (MIT; **32★**; HEAD `8ecdef6a6fcb`; + README SHA `0b24293bc84f`). Quote *theirs*: 485 catalog. + not affiliated with or endorsed by TypeSafe AI. + heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ + Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ + shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one. + catalog ≠ endorsement. + +4. **[NanmiCoder/jev-arena](https://github.com/NanmiCoder/jev-arena)** + - NEW HIGH measurement (MIT; **31★**; HEAD `2ca160cc4aa9`; + README SHA `4eb7f2dec20a`). Quote *theirs*: 10k comments. + 203.2s $0.84 vs 823.5s $1.50. 62.69% vs 67.26% *theirs* not gold. + strict 50.58% / 55.45%. AI-reviewed labels ≠ gold. *theirs* not Harbor. + Do **not** copy `npm`. + +5. **[openroboto-ai/jev-robot-control](https://github.com/openroboto-ai/jev-robot-control)** + - NEW HIGH measurement (NOASSERTION; **27★**; HEAD `7a4ed8b72c3c`; + README SHA `0e749c38197e`). Quote *theirs*: one seed-0 trial. + Jev $0.018825 vs Astra $5.93. one-trial robot ≠ Harbor. + game success ≠ calibrated Noul. + +6. **[endman100/research-Qwen3.8-JevLike](https://github.com/endman100/research-Qwen3.8-JevLike)** + - NEW HIGH measurement (license null; **0★**; HEAD `b406b17c3936`; + README SHA `7c908289d0e8`). Quote *theirs*: 10.59×. 6 class flips. + agreement ≠ accuracy. probabilities uncalibrated. Qwen3.8 ≠ Archer. + 10.59× systems ≠ ECE. *theirs* not Harbor. + +7. **[marcosmartinez/jev-acento](https://github.com/marcosmartinez/jev-acento)** + - NEW HIGH measurement (MIT; **0★**; HEAD `7e007b4c2bd5`; + README SHA `994943244b6b`). Quote *theirs*: Spanish −6.4 pp XNLI. + ECE 0.057→0.101. 72.2% vs 63.4% p_max≥0.9 coverage. + 19,200 calls, 3,200 paired, $0.58. *theirs* not Harbor. + +8. **[okinaaudio/live-jev](https://github.com/okinaaudio/live-jev)** + - NEW HIGH measurement (MIT; **36★**; HEAD `2446eb777ad9`; + README SHA `52e7d45c799f`). Quote *theirs*: Control Ableton Live + with one short sentence. Source-only; no packaged release. + serving substrate ≠ calibrated replica. Do **not** copy install recipes. + +9. **[win4r/jev-skill-suggester](https://github.com/win4r/jev-skill-suggester)** + - NEW HIGH measurement (MIT; **27★**; HEAD `05fbd7ce9ec7`; + README SHA `2a03bec83f3c`). Quote *theirs*: does not execute or + install candidate skills. local_only ≠ Jev. cutoff 0.80 / 0.65 + still soft. does not execute. + +10. **[PyModel/typesafe-mcp](https://github.com/PyModel/typesafe-mcp)** + - NEW HIGH measurement (MIT; **23★**; HEAD `cf01808bf274`; + README SHA `c7c36760aa1c`). Quote *theirs*: evaluate is a stdio + MCP server. The host still reasons, edits, and executes. + does not execute. serving substrate ≠ calibrated replica. + +11. **[Xubqpanda/JevLoop](https://github.com/Xubqpanda/JevLoop)** + - NEW HIGH measurement (Apache-2.0; **0★**; HEAD `50236bf2221b`; + README SHA `0e3f67c72801`). Quote *theirs*: 12:1 / 7.7% from a + rule table, not a model. rule-table ≠ model. *theirs* not Harbor. + +12. **[elberacasa/omawish](https://github.com/elberacasa/omawish)** + - NEW HIGH measurement (NOASSERTION; **0★**; HEAD `82c43e5425df`; + README SHA `bb2be3850163`). Quote *theirs*: 33M local; 60/66 + held-out; 0/124 wrong actions; 35ms. replica ≠ TypeSafe. + serving substrate ≠ calibrated replica. *theirs* not Harbor. + +13. **[Rizzo-AI-Academy/rizzo-flow](https://github.com/Rizzo-AI-Academy/rizzo-flow)** + - NEW HIGH measurement (Apache-2.0; **1★**; HEAD `d97ef676a40e`; + README SHA `529432bebd5f`). Quote *theirs*: local Jev-compatible; + ~250ms Q8. replica ≠ TypeSafe. serving substrate ≠ calibrated replica. + +### Remainder (short cards, same hour) + +- **[AkashPriyadarshii/jev-seo](https://github.com/AkashPriyadarshii/jev-seo)** : 100% free ₹0 agent-first SEO & GEO CLI suite and MCP server in Rust replacing Semrush and OpenSEO via DuckDuckGo and TypeSafe Jev System One (MIT; **21★**; HEAD `f8cb7c55c356`; README SHA `e3290fd15add`). *theirs*. catalog ≠ endorsement. +- **[devtooligan/jevscan-evm](https://github.com/devtooligan/jevscan-evm)** : (empty description) (MIT; **21★**; HEAD `2ecbe7cd93a9`; README SHA `075a01007b92`). *theirs*. catalog ≠ endorsement. +- **[mizzlelover/jev-hub](https://github.com/mizzlelover/jev-hub)** : JEV HUB · X 上关于 TypeSafe AI「系统一模型」Jev 的长文与演示视频聚合（保留原链与作者）｜ 谁是专家 出品 (NOASSERTION; **21★**; HEAD `4262d71d2d06`; README SHA `04530e25b751`). *theirs*. catalog ≠ endorsement. +- **[zszz3/Pi-Jev-Guide](https://github.com/zszz3/Pi-Jev-Guide)** : (empty description) (MIT; **19★**; HEAD `f187d464062b`; README SHA `45aaa093f82b`). *theirs*. catalog ≠ endorsement. +- **[clouatre-labs/decisions-judge-mcp](https://github.com/clouatre-labs/decisions-judge-mcp)** : MCP server exposing an LLM judge (typed decisions: yes/no probability, choice, score) for coding agents (Apache-2.0; **1★**; HEAD `b6fff8ee714c`; README SHA `cd771f8872c0`). *theirs*. catalog ≠ endorsement. +- **[sable-inc/jev-linter-action](https://github.com/sable-inc/jev-linter-action)** : Configurable semantic CI checks for repository files using TypeSafe Jev (MIT; **1★**; HEAD `1f6ba701fe72`; README SHA `b1ac97554b82`). *theirs*. catalog ≠ endorsement. +- **[shirenchuang/awsomejev](https://github.com/shirenchuang/awsomejev)** : Awesome Jev：Jev 开源生态导航 (MIT; **1★**; HEAD `a994099f0a32`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[1816586742-stack/jev-craft](https://github.com/1816586742-stack/jev-craft)** : 让 Agent 长出「操作杆」：用 System One 模型（Jev）承担高频判断、多模态模型当眼睛。给思路 + 可跑的参考实现（75 条离线断言，零依赖零 key） (MIT; **0★**; HEAD `ad4926f663b8`; README SHA `b614d4f0b313`). *theirs*. catalog ≠ endorsement. +- **[DolphinMiner/jev-rss](https://github.com/DolphinMiner/jev-rss)** : A local-first RSS reader with Jev-powered semantic screening. Follow what matters, inspect every judgment, and keep control of your reading. English / 简体中文. (MIT; **0★**; HEAD `4a2dbb770faa`; README SHA `caaf2f1ebb6a`). *theirs*. catalog ≠ endorsement. +- **[Dreydrey9000/jev-relay](https://github.com/Dreydrey9000/jev-relay)** : Guarded local-first decision advice for Claude Code, Codex and Hermes/Jax, with explicit Jev checks and review fallbacks. (MIT; **0★**; HEAD `d52ffe1fc023`; README SHA `2039013eb53d`). *theirs*. catalog ≠ endorsement. +- **[Eric-Zhou-0302/jev-A-share-trader](https://github.com/Eric-Zhou-0302/jev-A-share-trader)** : A Jev-powered technical analysis workspace for China A-shares, supporting AKShare/Tushare, market scans, and Buy/Hold/Sell assessments with time horizons and traceable evidence. (MIT; **0★**; HEAD `a7ad82306fbe`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[JTech-CO/Jev-Simulink-Supervisor](https://github.com/JTech-CO/Jev-Simulink-Supervisor)** : Jev-Simulink Supervisor (MIT; **0★**; HEAD `76bc14a89694`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[YuanKJing/Jev-as-Policy](https://github.com/YuanKJing/Jev-as-Policy)** : The highly anticipated open-source repository for JEV as Policy enables one-click setup of the simulation environment. Evaluations of Astra + JEV on benchmarks such as RoboTwin will also be released soon. (MIT; **0★**; HEAD `c2e1e17b6bb0`; README SHA `2b83b5db78bd`). *theirs*. catalog ≠ endorsement. +- **[Zafer-Liu/jev-demos](https://github.com/Zafer-Liu/jev-demos)** : (empty description) (license null; **0★**; HEAD `c875343d3780`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[aboisvert/jevvy](https://github.com/aboisvert/jevvy)** : Use jev model to augment csv files with inferred classification, scoring, or probability scores (Apache-2.0; **0★**; HEAD `14de871f7503`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[andrest04/jev-lab](https://github.com/andrest04/jev-lab)** : Local lab for learning and testing TypeSafe Jev (System One) (license null; **0★**; HEAD `3ef5eb16c288`; README SHA `741995e66b02`). *theirs*. catalog ≠ endorsement. +- **[andyrewlee/awesome-system-one](https://github.com/andyrewlee/awesome-system-one)** : Curated list of tools related to system one models (license null; **0★**; HEAD `1345d8d237d3`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[ashafizullah/jev-triage](https://github.com/ashafizullah/jev-triage)** : Automated issue & PR triage for open-source maintainers, powered by Jev (TypeSafe AI). (MIT; **0★**; HEAD `6907ccc444a2`; README SHA `96dcb0a58798`). *theirs*. catalog ≠ endorsement. +- **[blanket11/jev-guide-ja](https://github.com/blanket11/jev-guide-ja)** : jevの学習 (license null; **0★**; HEAD `d844446d8386`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[early-effect/hexis](https://github.com/early-effect/hexis)** : ZIO / Scala 3 SDK for TypeSafe System One (Jev) (Apache-2.0; **0★**; HEAD `c2d622c130ba`; README SHA `4ffae17a7060`). *theirs*. catalog ≠ endorsement. +- **[fruitymcdoo/JevChat](https://github.com/fruitymcdoo/JevChat)** : A chat interface built on Jev, TypeSafe's decision-only model: every word is a typed decision (license null; **0★**; HEAD `6e2bc022cbbe`; README SHA `abb03d83190a`). *theirs*. catalog ≠ endorsement. +- **[ismaelsoilet/jev-harness](https://github.com/ismaelsoilet/jev-harness)** : (empty description) (MIT; **0★**; HEAD `9686bfd363c7`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[javsanesq/jevlab](https://github.com/javsanesq/jevlab)** : A terminal workbench for learning, testing, and connecting TypeSafe Jev decisions (license null; **0★**; HEAD `df7198de1700`; README SHA `87d810d6a4b2`). *theirs*. catalog ≠ endorsement. +- **[joelakaufmann-lgtm/NRS-Navigator](https://github.com/joelakaufmann-lgtm/NRS-Navigator)** : Local Nevada statute search and a reproducible evaluation of Jev-assisted ranking against keyword search. (Apache-2.0; **0★**; HEAD `6c9394e6c6cb`; README SHA `dbe9da1004bd`). *theirs*. catalog ≠ endorsement. +- **[k-srkw/jev-playground](https://github.com/k-srkw/jev-playground)** : (empty description) (license null; **0★**; HEAD `d935d01d70ab`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[koteitan/laya-bot-det](https://github.com/koteitan/laya-bot-det)** : bot detector by laya for nostr (MIT; **0★**; HEAD `93120b7f808a`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[kzkhykw/jev-or-not](https://github.com/kzkhykw/jev-or-not)** : Jevる？モデル選択フローチャート (license null; **0★**; HEAD `916a83d68c9a`; README SHA `7cbe3a508fab`). *theirs*. catalog ≠ endorsement. +- **[lezgoverci/jev-docs](https://github.com/lezgoverci/jev-docs)** : (empty description) (license null; **0★**; HEAD `db9c9a85d63d`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[mednabouli/jev-ai-polymarket-copy-trading](https://github.com/mednabouli/jev-ai-polymarket-copy-trading)** : Automated Polymarket copy trading bot with MCP servers, Telegram alerts, and profitable wallet tracking. Zero API keys - uses Claude Code OAuth session auth. (MIT; **0★**; HEAD `aad77195fff5`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[promptgtm-shared/clay-jev-people-ranker](https://github.com/promptgtm-shared/clay-jev-people-ranker)** : Agent Skill and Python workflow for Clay lead scoring, B2B prospect qualification, and people-search ranking with TypeSafe JEV. (MIT; **0★**; HEAD `2494b65218d4`; README SHA `40402cc861cb`). *theirs*. catalog ≠ endorsement. +- **[sarathi-aiml/jevsql](https://github.com/sarathi-aiml/jevsql)** : Text-to-SQL where the model never writes SQL , typed, calibrated decisions (TypeSafe Jev) + code-assembled queries (license null; **0★**; HEAD `ba46b8b400a1`; README SHA `9c05b06bf11a`). *theirs*. catalog ≠ endorsement. +- **[theosunny/jev_stock](https://github.com/theosunny/jev_stock)** : (empty description) (MIT; **0★**; HEAD `0c0e3fbbb786`; README SHA `NO-SHA`). *theirs*. catalog ≠ endorsement. +- **[willgriffin/pi-fusion-matrix](https://github.com/willgriffin/pi-fusion-matrix)** : Multi-model deliberation for the pi coding agent: named fusions, per-slot fallback and routing, version-free aliases, conservative decision backends (license null; **0★**; HEAD `f5d2a1b89448`; README SHA `dfb8b07ad2c0`). *theirs*. catalog ≠ endorsement. + +- **[davila7/jev-explained](https://github.com/davila7/jev-explained)** : pedagogy; ~100ms; code thresholds (MIT; **14★**; HEAD `5cbe35e04609`; README SHA `7c69d941dc82`). *theirs*. catalog ≠ endorsement. +- **[yzfly/awesome-jev-zh](https://github.com/yzfly/awesome-jev-zh)** : Chinese catalog (CC0-1.0; **31★**; HEAD `7cce63fedfc5`; README SHA `f8c549471e6a`). yzfly/awesome-jev-zh ≠ heyjunpenn/awesome-jev. catalog ≠ endorsement. +- **[ckaraca/awesome-jev](https://github.com/ckaraca/awesome-jev)** : curated list (CC0-1.0; **7★**; HEAD `2207d19d91fa`; README SHA `3e58af61cceb`). ckaraca/awesome-jev ≠ heyjunpenn/awesome-jev. catalog ≠ endorsement. +- **[arunav25/jev-mcp](https://github.com/arunav25/jev-mcp)** : MCP + eval harness (MIT; **5★**; HEAD `036d32433c0d`; README SHA `a858bf2227da`). arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp. serving substrate ≠ calibrated replica. +- **[cobusgreyling/Jev](https://github.com/cobusgreyling/Jev)** : Unofficial TypeSafe Jev showcase (MIT; **3★**; HEAD `c636087edf41`; README SHA `9de601117711`). catalog ≠ endorsement. +- **[dinkarjuyal/jev-gepa](https://github.com/dinkarjuyal/jev-gepa)** : HF AlexWortega/openjev NLI into GEPA, not TypeSafe (license null; **0★**; HEAD `3d52ba9b2239`; README SHA `655920a0144a`). serving substrate ≠ calibrated replica. +- **[luckberonne/mini-jev](https://github.com/luckberonne/mini-jev)** : shell-command classifier (MIT; **0★**; HEAD `97d4b30b6f5d`; README SHA `795ab56ea245`). luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev. +- **[Kwwwww74/OpenJev](https://github.com/Kwwwww74/OpenJev)** : stub README (license null; **0★**; HEAD `423875b6088b`; README SHA `8f5ec58f92ea`). Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev. +- **[peach-zhang/typesafe-go](https://github.com/peach-zhang/typesafe-go)** : Go SDK (MIT; **0★**; HEAD `cd4b6bf028f7`; README SHA `109ec34a8247`). peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go. +- **[laidick/system-one-benchmark](https://github.com/laidick/system-one-benchmark)** : (license null; **0★**; HEAD `ccd4b79d1a51`; README SHA `5ffe72671473`). laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark. +- **[sahasrarjn/system-one](https://github.com/sahasrarjn/system-one)** : (license null; **0★**; HEAD `bb2383834d19`; README SHA NO-SHA). sahasrarjn/system-one ≠ sgoedecke/system-one. +- **[twilwa/pi-typesafe](https://github.com/twilwa/pi-typesafe)** : Pi extension (license null; **0★**; HEAD `3dc40b969dc3`; README SHA `195563911db2`). twilwa/pi-typesafe ≠ TheoOliveira/pi-jev. + +### Skips (thin / collision / name-match / sibling) + +- skip Zefan-Cai/Open-Jev densify open #53. +- skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54. +- ikermoel/open-alternative-jev already §49. +- nrdz-labs/fast-jev-opencode already §62. +- mallahyari/system-one-benchmark already §61. +- liao96312/jev-arena-nanojev namesake of NanoJev. +- emirbartu/opencode-system-one already listed. +- Empty default-branch / NO-SHA: RuipuCui/jev-harness, eteen12/jev-browser-automation, jacks3tr/Jev-Desktop, olivdx/jev-mcp, pattoor/JEV-agent-opencv. +- Krug2/JevLM-Open no README; caohy1988/jev-guard-smoke README 404 (smoke only vs leepokai/jev-guard). +- heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one. +- arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp. +- luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev. +- Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev. +- peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go. +- laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark. +- sahasrarjn/system-one ≠ sgoedecke/system-one. +- aboisvert/jevvy ≠ PanAchy/jevvy. +- andrest04/jev-lab ≠ javsanesq/jevlab. +- twilwa/pi-typesafe ≠ TheoOliveira/pi-jev. +- RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness. +- Archer rewrite: **promised_not_landed**. Hub archerhume/4rcherhume HTTP **401**. + +### Pulse (live REST this hour) + +Archer still promised_not_landed. Hub archerhume/4rcherhume HTTP **401**. +Jev-X-Sentiment-Analysis **136★**. live-jev **36★**. awesome-jev **32★**. +jev-arena **31★**. jev-robot-control **27★**. jev-skill-suggester **27★**. +llm-to-jev **3★** (desc rewrite; SHA unchanged). This hour does not +re-census SemIf / Laya likes / tracker; those numbers stay §119 until +a dedicated pulse. `invented_signal: false`. + +### Formal compose + anti-patterns + +Formal methods **compose** with scoring. A Noul is a SENSOR. A catalog +is an index. A dashboard Choice is not a fill. AI-reviewed labels are +not gold. One seed-0 trial is not Harbor. 10.59× is a systems timing, +not ECE. Agreement is not accuracy. A description rewrite is not a SHA +move. A rule-table demo is not a model. local_only is not Jev. A local +compatible API is not TypeSafe. Treating 62.69% as gold, $0.018825 as +Harbor, 10.59× as ECE, 485 as a grant, or Qwen3.8 as Archer is +soundness theater. does not execute. catalog ≠ endorsement. +AI-reviewed labels ≠ gold. one-trial robot ≠ Harbor. +10.59× systems ≠ ECE. agreement ≠ accuracy. +desc rewrite ≠ SHA/behavior change. SHA move is not a replica. +Qwen3.8 ≠ Archer. + +### Adversarial review + testing hooks (Basit standing +order) + +Parent merge only after **CLEAN** adversarial review **AND** testing. +Hooks for the reviewer: + +- Uniqueness-gate: the consecutive `Hourly 1946 uniqueness lock:` + string must appear in every overlay listed below. Prior walls 0843 / + 0915 / jcr / 0922 / 0940 / 0947 / 1049 / 1143 / 1248 / 1340 / 1441 / + 1542 / 1643 / 1746 / 1843 stay one substring each (do not mutate them; + do not reopen #23–#52). §130 / 497–504 / #112 stay unused. +- Namesake locks: heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ + MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ + yzfly/awesome-jev-zh; arunav25/jev-mcp ≠ jkudish/jev-mcp; + luckberonne/mini-jev ≠ r-ms/mini-jev; Kwwwww74/OpenJev ≠ + razorback16/openjev; sahasrarjn/system-one ≠ sgoedecke/system-one. +- Densify vs new: llm-to-jev densifies §118. Open-Jev stays #53. + sgoedecke / mithalouni / kotoba stay #54. Do not mint sibling + first-sighting sections for those. +- Harbor-jevals: 62.69% / 203.2s $0.84 / $0.018825 / 10.59× / + −6.4 pp / ECE 0.057→0.101 / 72.2% vs 63.4% are *theirs*, not Harbor. +- Anti-patterns to refuse: TypeSafe drop-in; catalog as endorsement; + AI review as gold; one trial as Harbor; 10.59× as ECE; agreement as + accuracy; desc rewrite as SHA change; rule table as model; + Qwen3.8 as Archer; copying keys / `npm` / `pip` / `uv` / `docker`. +- Overlay set: SKILL.md body (protocol fragments + class-table densify + + Hourly 1946), mental-models Apply 1946, composition-algebra items + 505–520, faq, mixed-architecture, validation, toolbox-mapping, + methods-catalog, formal-methods, formal-semi-formal, + applied-mappings, judgment-class, question-design, + agent-self-assessment, mappings, CHANGELOG, README, docs/ecosystem, + findings batch #113, refresh-log, changelog-hourly.md, + revisit_fingerprints.json (llm-to-jev desc_hash; novel HIGH seeds). +- Offline check: `evaluate_decisions.py --self-test` (now includes + does not execute / AI-reviewed ≠ gold / one-trial ≠ Harbor / + 10.59× ≠ ECE / agreement ≠ accuracy / desc rewrite ≠ SHA) and + `uniqueness_gate.py` (0843 + 0915 + jcr / 0922 / 0940 / 0947 / + 1049 / 1143 / 1248 / 1340 / 1441 / 1542 / 1643 / 1746 / 1843 / 1946). + No live Jev key. No wrappers. + +Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 diff --git a/research/refresh-log.md b/research/refresh-log.md index 1a9af3d..b3e2fef 100644 --- a/research/refresh-log.md +++ b/research/refresh-log.md @@ -1,3 +1,20 @@ +## 2026-09-21 ~01:46 UTC / ~19:46 Boise - Hourly 1946 HIGH +- Fresh PR off latest `main` after merged #54 (aisearchio / `notes.md` §130 + / items 497–504 / batch #112) and merged #52 (hourly 1843 / `notes.md` §129 + / items 481–496 / batch #111). This fold: `notes.md` §131 / composition + 505–520 / findings batch #113. Do not reclaim §130 / 497–504 / #112. + **HARD RULE:** do not reopen or amend PR #23–#52. +- PRIMARY: X-sentiment does not execute. heyjunpenn 485 catalog. + jev-arena AI-reviewed ≠ gold. one seed-0 robot trial. 10.59× uncalibrated. + Spanish −6.4 pp. REVISIT: llm-to-jev description rewrite SHA unchanged. + Skip Open-Jev #53. Skip #54 three. +- Evidence: `research/archive/hourly/2026-09-21T01/`. +- uniqueness_gate 0843+0915+jcr+0922+0940+0947+1049+1143+1248+1340+1441+1542+1643+1746+1843+1936+1946. + Evaluator: does not execute / AI-reviewed ≠ gold / one-trial ≠ Harbor / + 10.59× ≠ ECE / agreement ≠ accuracy / desc rewrite ≠ SHA. + Quote *theirs*. No wrappers. `invented_signal: false`. +- Hourly 1946 uniqueness lock: brainstormity/Jev-X-Sentiment-Analysis 136★ HEAD 5c932f941a92 README SHA bf4134b44cda; platform does not execute trades; heyjunpenn/awesome-jev 485 catalog ≠ endorsement; heyjunpenn/awesome-jev ≠ yibie/awesome-jev ≠ MrJev/awesome-jev ≠ Promethe-us/awesome-jev ≠ ckaraca/awesome-jev ≠ yzfly/awesome-jev-zh ≠ shirenchuang/awsomejev ≠ andyrewlee/awesome-system-one; NanmiCoder/jev-arena 10k comments 62.69% vs 67.26% *theirs* not gold; 203.2s $0.84 vs 823.5s $1.50 *theirs*; AI-reviewed labels ≠ gold; openroboto-ai/jev-robot-control one seed-0 trial *theirs*; Jev $0.018825 vs Astra $5.93 *theirs*; one-trial robot ≠ Harbor; endman100/research-Qwen3.8-JevLike 10.59× *theirs*; 6 class flips; agreement ≠ accuracy; probabilities uncalibrated; Qwen3.8 ≠ Archer; 10.59× systems ≠ ECE; marcosmartinez/jev-acento Spanish −6.4 pp XNLI *theirs*; ECE 0.057→0.101 *theirs*; 72.2% vs 63.4% p_max≥0.9 coverage *theirs*; alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts; SHA unchanged 234058ab372d; 3★; heuristic conversion ≠ calibrated Noul; desc rewrite ≠ SHA/behavior change; skip Zefan-Cai/Open-Jev densify open #53; skip sgoedecke/system-one mithalouni/system-one-open kotoba-lang/typed-decisions open #54; ikermoel/open-alternative-jev already §49; nrdz-labs/fast-jev-opencode already §62; mallahyari/system-one-benchmark already §61; does not execute; catalog ≠ endorsement; *theirs* not Harbor; SHA move is not a replica; local_only ≠ Jev; rule-table ≠ model; replica ≠ TypeSafe; arunav25/jev-mcp ≠ jkudish/jev-mcp ≠ ThePFMind/jev-mcp ≠ burnigtm/jev-mcp; luckberonne/mini-jev ≠ r-ms/mini-jev ≠ samatv256/mini-Jev; Kwwwww74/OpenJev ≠ razorback16/openjev ≠ kyegomez/open-jev ≠ Zefan-Cai/Open-Jev; peach-zhang/typesafe-go ≠ kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go; laidick/system-one-benchmark ≠ mallahyari/system-one-benchmark; sahasrarjn/system-one ≠ sgoedecke/system-one; aboisvert/jevvy ≠ PanAchy/jevvy; andrest04/jev-lab ≠ javsanesq/jevlab; twilwa/pi-typesafe ≠ TheoOliveira/pi-jev; RuipuCui/jev-harness ≠ ismaelsoilet/jev-harness ≠ AntonioCoppe/jev-harness; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §131 + ## 2026-09-21T01:36Z / ~19:36 Boise - User-provided 1936 HIGH - Fresh PR off latest `main` after merged #52 (hourly 1843 / `notes.md` §129 / items 481–496 / batch #111) and merged #51 (hourly 1746 / `notes.md` @@ -30,7 +47,6 @@ Quote *theirs*. No wrappers. `invented_signal: false`. - User-provided Open-Jev densify uniqueness lock: Zefan-Cai/Open-Jev densify HEAD 4933ee84951f README SHA ce1a587219e4; pushed 2026-09-21T01:34Z; Astra TREC commit 1dd56990be7e pushed 2026-09-21T01:17Z; live 3★ (was 0★; star-noise is not the fold); LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48; 27B still in progress; Independent of TypeSafe; no RLCD/parity claims; LoRA ≠ RLCD replica; customer-service P50 local HTTP 85.03 ms vs Jev HTTPS 295.26 ms *theirs*; 1024 tokens/32 candidates Open-Jev slower 1015.90 vs 301.37 *theirs*; prefix caching experimental/off by default; CUDA prefix caching exceeded tolerance on 9/11 workloads; systems latency ≠ semantic equivalence; GPT Luna P50 918.13 ms Astra 1938.39 ms *theirs*; TREC-DL Jev/Luna/Astra completed; Open-Jev TREC pending; 80,816 training rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97% *theirs* not Harbor; hard acc ≠ calibrated Noul; type-valid ≠ exact; Qwen/Qwen3.8-27B ≠ Archer; website https://zefan-cai.github.io/open-jev/; launch X thread https://x.com/Zefan_Cai/status/2101782158658695388 https://x.com/Zefan_Cai/status/2101786019607740436 https://x.com/Zefan_Cai/status/2101789698947793231; densify §125 not a sibling first sighting; SHA move is not a replica; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43/#44/#45/#46/#47/#48/#49/#50/#51/#52; notes.md §125 - ## 2026-09-21 ~00:43 UTC / ~18:43 Boise - Hourly 1843 HIGH - Fresh PR off latest `main` after merged #51 (hourly 1746 / `notes.md` §128 / items 465–480 / batch #110) and merged #50 (hourly 1643 / `notes.md` diff --git a/research/revisit_fingerprints.json b/research/revisit_fingerprints.json index 15c9c8a..1cb1974 100644 --- a/research/revisit_fingerprints.json +++ b/research/revisit_fingerprints.json @@ -20,7 +20,7 @@ "forks_count", "likes" ], - "note": "Last-look snapshots for already-catalogued sources. Hourly diffs these four fingerprints. Star-noise is not a fold. SHA move is not a replica. Seeded from merged notes (SemIf \u00a7117, NanoJev \u00a7115) plus hourly 1248 densify (openjev \u00a775, von \u00a749, verdict \u00a771, kev \u00a745, jeff \u00a760, TypeLLM \u00a7113) plus hourly 1340 densify (openjev sdk 0.7 + MLX 400, kev PLAN_Qwen35) plus hourly 1441 densify (openjev STE backends+Codiv dual serving) and 1441 first sightings plus hourly 1542 densify (TypeLLM README 3k\u219212k B, kev family new-source) and 1542 first sightings (pi-jev densify, jev-sentinel, tool-routers, leanest, jevals, MrJev, jev-firewall, jev-codex-approval, HF encoder/quanto). default_sha is the full HEAD. pushed_at is the GitHub push clock. description_hash is sha256[:12] of the GitHub/Space description, or null when notes do not quote it. README SHA is an optional extra, not a substitute for description_hash. plus hourly 1746 densify (TypeLLM truncated thinking + qwen35_small, kev 0.8B Qwen3.5 family, jev-pruner \u00a753, jevassert \u00a770, jev-packs \u00a764) and 1746 first sightings (vexjoy, Canny, jev-engineering, five-lines, jev-table, kev-ane) plus hourly 1843 densify (kev own-data JSONL / --init_from warm-start, SHA 8465c4c4c294 \u2192 bd058057ad0a) and 1843 first sightings (assay-001, kyegomez reconstruction, jev-ra, catalogs). plus user-provided 1936 first sightings (sgoedecke/system-one, mithalouni/system-one-open, kotoba-lang/typed-decisions). plus Open-Jev densify 2026-09-21 (github:Zefan-Cai/Open-Jev SHA 6d8de5ed72a0 → 4933ee84951f; README 771bf3135f50 → ce1a587219e4; HF 2B/9B packs + dataset last_look; densify §125 not a sibling).", +"note": "Last-look snapshots for already-catalogued sources. Hourly diffs these four fingerprints. Star-noise is not a fold. SHA move is not a replica. Seeded from merged notes (SemIf §117, NanoJev §115) plus hourly 1248 densify (openjev §75, von §49, verdict §71, kev §45, jeff §60, TypeLLM §113) plus hourly 1340 densify (openjev sdk 0.7 + MLX 400, kev PLAN_Qwen35) plus hourly 1441 densify (openjev STE backends+Codiv dual serving) and 1441 first sightings plus hourly 1542 densify (TypeLLM README 3k→12k B, kev family new-source) and 1542 first sightings (pi-jev densify, jev-sentinel, tool-routers, leanest, jevals, MrJev, jev-firewall, jev-codex-approval, HF encoder/quanto). default_sha is the full HEAD. pushed_at is the GitHub push clock. description_hash is sha256[:12] of the GitHub/Space description, or null when notes do not quote it. README SHA is an optional extra, not a substitute for description_hash. plus hourly 1746 densify (TypeLLM truncated thinking + qwen35_small, kev 0.8B Qwen3.5 family, jev-pruner §53, jevassert §70, jev-packs §64) and 1746 first sightings (vexjoy, Canny, jev-engineering, five-lines, jev-table, kev-ane) plus hourly 1843 densify (kev own-data JSONL / --init_from warm-start, SHA 8465c4c4c294 → bd058057ad0a) and 1843 first sightings (assay-001, kyegomez reconstruction, jev-ra, catalogs). plus user-provided 1936 first sightings (sgoedecke/system-one, mithalouni/system-one-open, kotoba-lang/typed-decisions). plus Open-Jev densify 2026-09-21 (github:Zefan-Cai/Open-Jev SHA 6d8de5ed72a0 → 4933ee84951f; README 771bf3135f50 → ce1a587219e4; HF 2B/9B packs + dataset last_look; densify §125 not a sibling) plus hourly 1946 densify (alexwestco/llm-to-jev description rewrite Convert LLM prompts to Jev prompts, SHA unchanged 234058ab372d, 3★) and 1946 first sightings (X-sentiment, awesome-jev 485, jev-arena, robot-control, Qwen3.8-JevLike, jev-acento).", "looks": [ { "id": "github:TheoLeeCJ/SemIf", @@ -1590,6 +1590,162 @@ "release_tag": null }, "readme_sha": "4d6bbf4c4e446270dea99bb8b9d70bf2df628f09" + }, + { + "id": "github:alexwestco/llm-to-jev", + "notes_section": "118", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "234058ab372d7754833c6279601755f8fda55d98", + "pushed_at": "2026-09-20T12:06:01Z", + "description_hash": "9f521cf7cc20", + "release_tag": null + }, + "readme_sha": "43cd94fba7527d78635b148628485c0cd1b1ad66" + }, + { + "id": "github:brainstormity/Jev-X-Sentiment-Analysis", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "5c932f941a92348781e8e5b471a9ddb6af980253", + "pushed_at": "2026-09-19T22:51:11Z", + "description_hash": null, + "release_tag": null + }, + "readme_sha": "bf4134b44cda73599f005e9408391353a2ed9437" + }, + { + "id": "github:heyjunpenn/awesome-jev", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "8ecdef6a6fcb3acc0c498fe8a87855e5bcd21c0f", + "pushed_at": "2026-09-20T10:08:27Z", + "description_hash": "eed807328057", + "release_tag": null + }, + "readme_sha": "0b24293bc84ff0cc828b67c7d5049b74d0d085ed" + }, + { + "id": "github:NanmiCoder/jev-arena", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "2ca160cc4aa9ac72a4341e2ac5903258e8c69c84", + "pushed_at": "2026-09-20T15:00:39Z", + "description_hash": "e2732659dccb", + "release_tag": null + }, + "readme_sha": "4eb7f2dec20a2ecdf0d74a6254f311ee8bff19a7" + }, + { + "id": "github:openroboto-ai/jev-robot-control", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "7a4ed8b72c3c17d7aa790678ed9660df67c10dd3", + "pushed_at": "2026-09-19T13:39:07Z", + "description_hash": null, + "release_tag": null + }, + "readme_sha": "0e749c38197e78c46dacbb1f61910ef2b244a823" + }, + { + "id": "github:endman100/research-Qwen3.8-JevLike", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "b406b17c3936e23ba588943cf05a15b8b27e0943", + "pushed_at": "2026-09-21T01:31:38Z", + "description_hash": "90a2bc4aa36a", + "release_tag": null + }, + "readme_sha": "7c908289d0e8a5ba36dbef64034ecf281dd2d89a" + }, + { + "id": "github:marcosmartinez/jev-acento", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "7e007b4c2bd552471b56eff035f9a3df7593f9fe", + "pushed_at": "2026-09-21T01:12:19Z", + "description_hash": "1f8cc7248824", + "release_tag": null + }, + "readme_sha": "994943244b6ba1fdaf3420b2dcf5a9066f158509" + }, + { + "id": "github:okinaaudio/live-jev", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "2446eb777ad9f59f77b96ee5b081ee8c2812e0a3", + "pushed_at": "2026-09-21T01:46:35Z", + "description_hash": "73f6f5c579e5", + "release_tag": null + }, + "readme_sha": "52e7d45c799f9b4d20053d2c7e28b8caac0e1b07" + }, + { + "id": "github:win4r/jev-skill-suggester", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "05fbd7ce9ec74a2f193c09276a8d4d077d9d5e5c", + "pushed_at": "2026-09-19T15:31:40Z", + "description_hash": "a146a470c049", + "release_tag": null + }, + "readme_sha": "2a03bec83f3c9915d7c8691523a7a489c65fee7e" + }, + { + "id": "github:PyModel/typesafe-mcp", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "cf01808bf27485a017cc6eebead6384556dab552", + "pushed_at": "2026-09-21T01:48:45Z", + "description_hash": null, + "release_tag": null + }, + "readme_sha": "c7c36760aa1cfb1c4db3bafdc867b4bafd1dd241" + }, + { + "id": "github:Xubqpanda/JevLoop", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "50236bf2221bdc9b3f556ad4f6e067daf558c12b", + "pushed_at": "2026-09-21T01:46:46Z", + "description_hash": "ab62fbd74fcf", + "release_tag": null + }, + "readme_sha": "0e3f67c72801630bee7149c7e8157ea1370c37fa" + }, + { + "id": "github:elberacasa/omawish", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "82c43e5425df4acf75459a9350115d0a849c5f0b", + "pushed_at": "2026-09-21T01:48:04Z", + "description_hash": "e53948cca959", + "release_tag": null + }, + "readme_sha": "bb2be38501638237b1b16ea70f1302b48b78861d" + }, + { + "id": "github:Rizzo-AI-Academy/rizzo-flow", + "notes_section": "131", + "last_look": "2026-09-21T01:46Z", + "fingerprints": { + "default_sha": "d97ef676a40e2ad47541835f8676b4be60331911", + "pushed_at": "2026-09-21T01:55:02Z", + "description_hash": "d006ffab3676", + "release_tag": null + }, + "readme_sha": "529432bebd5f6692f1ccffe6bbb0176dbc446aa9" } ] } From d5bb6b8679e18bbba639d6f902671bb58c2c2b1d Mon Sep 17 00:00:00 2001 From: Cursor Agent Date: Mon, 21 Sep 2026 02:50:32 +0000 Subject: [PATCH 2/2] =?UTF-8?q?Lock=20llm-to-jev=20densify=20to=20notes.md?= =?UTF-8?q?=20=C2=A7118?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit densify_original_ids was missing alexwestco/llm-to-jev, so moving the store notes_section off 118 to 131 would have passed the revisit self-test. Same densify-not-sibling gate as Open-Jev on §125. Does not bump 0.5.0. Co-authored-by: Basit Mustafa <24601@users.noreply.github.com> --- research/revisit_fingerprints.py | 1 + 1 file changed, 1 insertion(+) diff --git a/research/revisit_fingerprints.py b/research/revisit_fingerprints.py index 2db7508..9682c64 100644 --- a/research/revisit_fingerprints.py +++ b/research/revisit_fingerprints.py @@ -302,6 +302,7 @@ def self_test() -> None: "hf:ZefanCai/Open-Jev-2B": "125", "hf:ZefanCai/Open-Jev-9B": "125", "hf:ds:ZefanCai/Open-Jev": "125", + "github:alexwestco/llm-to-jev": "118", } for look_id, section in densify_original_ids.items(): assert look_id in by_id, look_id
\"Typing
As you type: apps, windows, commands, arithmetic, earlier wishes. Instant, offline.		\"'what
A question is answered in the bar. Enter copies it.
\"'remind
Values come from your own words: the number, and the phrase.		\"'update
Unsure, it asks. Tab shows what was weighed.