From c42bbfc1a7b48fe7154b01d9c8e00b806c83a407 Mon Sep 17 00:00:00 2001 From: Cursor Agent Date: Mon, 21 Sep 2026 11:08:11 +0000 Subject: [PATCH 1/2] Fold hourly 0445 HIGH MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Catalog the 0445 Boise watch into notes §140 / composition 649–664 / findings batch #122. Lock lcc keep-all vs withdrawn mock, any2jev ECE 0.027 *theirs*, Jev-Compatible softmax gateway, NanoJev JevHarness densify §115, and third-party arcade/survey benches as *theirs*. Co-authored-by: Basit Mustafa <24601@users.noreply.github.com> --- .agents/skills/augustus/SKILL.md | 18 +- .../references/agent-self-assessment.md | 1 + .../augustus/references/applied-mappings.md | 1 + .../references/composition-algebra.md | 75 + .agents/skills/augustus/references/faq.md | 1 + .../augustus/references/formal-methods.md | 1 + .../augustus/references/formal-semi-formal.md | 1 + .../augustus/references/judgment-class.md | 1 + .../skills/augustus/references/mappings.md | 1 + .../augustus/references/mental-models.md | 32 + .../augustus/references/methods-catalog.md | 1 + .../augustus/references/mixed-architecture.md | 1 + .../augustus/references/question-design.md | 1 + .../augustus/references/toolbox-mapping.md | 1 + .../skills/augustus/references/validation.md | 1 + .../augustus/scripts/evaluate_decisions.py | 54 + .../augustus/scripts/uniqueness_gate.py | 80 +- CHANGELOG.md | 37 + README.md | 2 + docs/ecosystem.md | 3 + research/archive/findings.md | 28 + .../hourly/2026-09-21T11/augustus_items.json | 812 +++++++ .../archive/hourly/2026-09-21T11/github.json | 1917 +++++++++++++++++ research/archive/hourly/2026-09-21T11/hf.json | 415 ++++ .../2026-09-21T11/novel_high_this_run.json | 738 +++++++ .../hourly/2026-09-21T11/prompt_Augustus.md | 71 + .../2026-09-21T11/revisit_high_this_run.json | 76 + .../hourly/2026-09-21T11/run_digest.json | 11 + research/changelog-hourly.md | 14 + research/notes.md | 239 ++ research/refresh-log.md | 18 + research/revisit_fingerprints.json | 240 ++- research/revisit_fingerprints.py | 17 +- 33 files changed, 4892 insertions(+), 17 deletions(-) create mode 100644 research/archive/hourly/2026-09-21T11/augustus_items.json create mode 100644 research/archive/hourly/2026-09-21T11/github.json create mode 100644 research/archive/hourly/2026-09-21T11/hf.json create mode 100644 research/archive/hourly/2026-09-21T11/novel_high_this_run.json create mode 100644 research/archive/hourly/2026-09-21T11/prompt_Augustus.md create mode 100644 research/archive/hourly/2026-09-21T11/revisit_high_this_run.json create mode 100644 research/archive/hourly/2026-09-21T11/run_digest.json diff --git a/.agents/skills/augustus/SKILL.md b/.agents/skills/augustus/SKILL.md index 2112d89..a48b92b 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", 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 / jaredpalmer/kev densify HEAD c096660c8da2 / PLAN SHA 8d77dd271c66 / README SHA unchanged 84b872488915 / night-2 dates/unknowable/assertion / KEV_TEMPERATURE T≈2.0 / Brier 0.291→0.267 ECE 0.105→0.039 *theirs* / 7.5%→3.2% *theirs* / grouped T rejected / Qwen3.6-35B-A3B smoke 0.812 *theirs* / 21M LoRA experts frozen / Hub --revision night2-du / MMLU-Pro 1000 Kev-9B 0.511 Jev 0.829 *theirs* / Qwen3.6 ≠ Archer / temperature scaling ≠ ECE unless measured / Hub --revision is a pin not a replica / kotoba-lang/typed-decisions densify HEAD ff7f84e74d04 / feat expose trained OpenJev decision runtime / open_jev.py / tests/test_open_jev.py / generated_text: False / trained runtime ≠ TypeSafe / OpenJev.from_pretrained / decide_request kind typed-decisions/open-jev-v1 / daftAI2026/awesome-jev ≠ heyjunpenn/awesome-jev / franckverrot/lev ≠ jaredpalmer/kev / neko233-com/laya-go ≠ convaiinnovations/laya / tryAGI/TypeSafeAI ≠ official / jaanavit/gliner2-skill Locate ≠ decide / hourly 2049 / notes.md §132 / Open-Jev densify HEAD a00559ea0ab2 / README SHA unchanged ce1a587219e4 / Publish prepared Open-Jev provider quality evaluation pipeline / 808 requests 1841 labelled decisions per model / Open-Jev GPU inference has not started / 48 CPU tests pass / Open-Jev TREC pending / 65/76 72/76 66/76 60/76 71/76 *theirs* / provider pipeline ≠ completed Open-Jev quality / CPU tests ≠ GPU scores / tinmanlab/cartpole-jev densify HEAD 922cc61490a0 / Active model Kev Not TypeSafe Jev / 81.25% 52/64 *theirs* / one record of 64 / fine-tuned Kev ≠ TypeSafe Jev / softmax ≠ calibrated Noul / xuboboo/ashare-trader densify HEAD 26c7e95e6828 / QMT sidecar mock/dry default no orders / AUC 0.532 *theirs* / does not execute / gauravsaini/kevin first card Playwright + Onyx / 3.69ms *theirs* not Harbor / metask-jev-4b 79.6% / 80.1% *theirs* / cutoff 95% still soft / hourly 2146 / notes.md §133 / Open-Jev densify HEAD 48346d0630f1 / Publish strict Open-Jev TREC evaluation preparation and context proof / Actual Open-Jev TREC model inference is pending / All 79 combined CPU tests pass / TREC prep ≠ completed Open-Jev TREC / context proof ≠ nDCG / TypeLLM/TypeLLM densify HEAD 8a8b4aefd443 / typellm 0.1.1 / PyPI packaging ≠ calibrated Noul / featherless-ai/simple-jev 408★ HEAD b02aa81c915a / logits are not calibrated probabilities of correctness / hourly 2246 / notes.md §134 / razorback16/openjev densify HEAD 2050fdb8280d / MLX backend steps>1/think/text gen + image Qs / dual serving is not generate / Hosted Codiv ≠ TypeSafe / TypeLLM/TypeLLM densify HEAD 8a8b4aefd443 / GitHub Release v0.1.1 / README SHA unchanged 9f6dea3a4c8c / Constrained AR ≠ calibrated Noul / PyPI packaging ≠ calibrated Noul / zjunlp/JevLoop 6★ independent not affiliated / WANLI-256 74.6% *theirs* / option order 0.188 or 0.542 *theirs* / LabGuy94/jevtok 0 mismatches *theirs* not Harbor / ockev 35ms 95.8% TomatoEggBench *theirs* not Harbor / n=8 is not Harbor / ranking before lossless condensation / llm-routing-jiv does not execute / skip-thin layacm empty SHA / hourly 2347 / notes.md §135 / Open-Jev densify HEAD f46ff604f794 / README SHA e32c4bbd519c / Publish audited JevBench public-subset baselines / public-subset ≠ Harbor / 231 ≠ 534 / jaredpalmer/kev densify HEAD e0bcf50153f1 / PLAN correct 35B MMLU-Pro (0.550) / evaluate.load honour weights_dtype=bf16 / wy-coliney/jev-browser-use 282★ / 5-10× *theirs* not Harbor / fail-open routing ≠ permission / ordered routing ≠ end-to-end / softmax next-token ≠ calibrated Noul / potential_match ≠ hiring decision / hourly 0049 / notes.md §136 / Open-Jev densify HEAD ed45657bf726 / README SHA 12e0f581e15d / Publish audited v3 community data and held-out evaluation protocol / Redesign readable project site and consolidate benchmark results / v3 data prepared ≠ retrained released models / held-out protocol ≠ Harbor / 1,280-row panel ≠ Harbor / finite training loss ≠ quality improvement / website redesign ≠ calibration / 129,288 decision rows 74,921 training / frozen mixture 96,849 training / 1,280-row / 840-group comparison panel / 27B step 616 pending / naive throws away 83% *theirs* / certo KL 0.008 acc 0.844 ECE 0.004 *theirs* / first-instinct 63.3%→78.1% *theirs* not Harbor / 371,278 prepared ≠ consumed / Jev is a gate not a generator / Lake remains admission / Jev never writes Lean / community port ≠ TypeSafe / chy4pro/jev-for-chrome ≠ browser-use/jev-ultrafast / joint RLCD *theirs* / Dohnuts ≠ TypeSafe / Akashdb5/jev-router ≠ gargpratyush/jev-router ≠ daviddl9/jev-router / kiuckhuang/laya-jev ≠ KonghaYao/laya-jev / tegersdorfer-collab/jevkit ≠ isiomaC/jevkit ≠ WaynezProg/jev-kit / buluoray/JevOnly already carded / yottayoshida/jev-intent-review already carded / skip-thin Iskandeur/system1-system2 zhlei07/open-system-one Hand-In/openjev-multimodal gwxcsny53/jev-watchtower empty SHA / hourly 0151 / notes.md §137 / GLiClass knowledgator Hub family class-peer catalog not Jev equivalent / typed-decision-leaderboard *theirs* not Harbor / JEV 0.7350 ZTC 27B 0.7289 / Jevbridge ACP and MCP adapter / Cut the slop / Not a Cua binding / open reproductions of the shape / 82.3% ECE 0.017 *theirs* / hourly 0248 / notes.md §138, or "cascade sign-flip / calibration theater": read `references/faq.md`, / Greedy 0.90 vs Oracle 0.82 *theirs* / Random conf 0.00 still 20.5% *theirs* / 8,400 calls $0.39 *theirs* / Noul 0.7 true 44% *theirs* / JevBench 81.65 *theirs* not Harbor / WindTunnel 49/49 *theirs* not Harbor / 0-byte Mandelbrot is not a replica / training not complete / Compose meaning like state / Code enumerates the candidates / Jev is the first classifier the design is bound to none / context is the conversation so far / A clean report is not proof / does not sandbox / ChatJEVs ≠ erik-dunteman/ChatJev / generation from Choice is not a language model replica / demo scores are not accuracy measurements / Qwen2.5 ≠ Archer / confidence ≠ P(correct) / seed 42 n=1 is not Harbor / meijustory123/OpenJev-Kit IS meijustory123/openjev / skip-thin Fibonaccirabbit/Jev-GalGame MadhavBahl/jev-guide advance-lion/dsh-jev-hooks amithgc/local-jev hiro1202/jev-review-gate-poc inlight37-design/decision-model_lab kuhung/ask-jev mmiguez314/jev-lab pomodorozhong/exp-jev vanthiet1/JevGuarAgent empty SHA / hourly 0348 / notes.md §139 + 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 / jaredpalmer/kev densify HEAD c096660c8da2 / PLAN SHA 8d77dd271c66 / README SHA unchanged 84b872488915 / night-2 dates/unknowable/assertion / KEV_TEMPERATURE T≈2.0 / Brier 0.291→0.267 ECE 0.105→0.039 *theirs* / 7.5%→3.2% *theirs* / grouped T rejected / Qwen3.6-35B-A3B smoke 0.812 *theirs* / 21M LoRA experts frozen / Hub --revision night2-du / MMLU-Pro 1000 Kev-9B 0.511 Jev 0.829 *theirs* / Qwen3.6 ≠ Archer / temperature scaling ≠ ECE unless measured / Hub --revision is a pin not a replica / kotoba-lang/typed-decisions densify HEAD ff7f84e74d04 / feat expose trained OpenJev decision runtime / open_jev.py / tests/test_open_jev.py / generated_text: False / trained runtime ≠ TypeSafe / OpenJev.from_pretrained / decide_request kind typed-decisions/open-jev-v1 / daftAI2026/awesome-jev ≠ heyjunpenn/awesome-jev / franckverrot/lev ≠ jaredpalmer/kev / neko233-com/laya-go ≠ convaiinnovations/laya / tryAGI/TypeSafeAI ≠ official / jaanavit/gliner2-skill Locate ≠ decide / hourly 2049 / notes.md §132 / Open-Jev densify HEAD a00559ea0ab2 / README SHA unchanged ce1a587219e4 / Publish prepared Open-Jev provider quality evaluation pipeline / 808 requests 1841 labelled decisions per model / Open-Jev GPU inference has not started / 48 CPU tests pass / Open-Jev TREC pending / 65/76 72/76 66/76 60/76 71/76 *theirs* / provider pipeline ≠ completed Open-Jev quality / CPU tests ≠ GPU scores / tinmanlab/cartpole-jev densify HEAD 922cc61490a0 / Active model Kev Not TypeSafe Jev / 81.25% 52/64 *theirs* / one record of 64 / fine-tuned Kev ≠ TypeSafe Jev / softmax ≠ calibrated Noul / xuboboo/ashare-trader densify HEAD 26c7e95e6828 / QMT sidecar mock/dry default no orders / AUC 0.532 *theirs* / does not execute / gauravsaini/kevin first card Playwright + Onyx / 3.69ms *theirs* not Harbor / metask-jev-4b 79.6% / 80.1% *theirs* / cutoff 95% still soft / hourly 2146 / notes.md §133 / Open-Jev densify HEAD 48346d0630f1 / Publish strict Open-Jev TREC evaluation preparation and context proof / Actual Open-Jev TREC model inference is pending / All 79 combined CPU tests pass / TREC prep ≠ completed Open-Jev TREC / context proof ≠ nDCG / TypeLLM/TypeLLM densify HEAD 8a8b4aefd443 / typellm 0.1.1 / PyPI packaging ≠ calibrated Noul / featherless-ai/simple-jev 408★ HEAD b02aa81c915a / logits are not calibrated probabilities of correctness / hourly 2246 / notes.md §134 / razorback16/openjev densify HEAD 2050fdb8280d / MLX backend steps>1/think/text gen + image Qs / dual serving is not generate / Hosted Codiv ≠ TypeSafe / TypeLLM/TypeLLM densify HEAD 8a8b4aefd443 / GitHub Release v0.1.1 / README SHA unchanged 9f6dea3a4c8c / Constrained AR ≠ calibrated Noul / PyPI packaging ≠ calibrated Noul / zjunlp/JevLoop 6★ independent not affiliated / WANLI-256 74.6% *theirs* / option order 0.188 or 0.542 *theirs* / LabGuy94/jevtok 0 mismatches *theirs* not Harbor / ockev 35ms 95.8% TomatoEggBench *theirs* not Harbor / n=8 is not Harbor / ranking before lossless condensation / llm-routing-jiv does not execute / skip-thin layacm empty SHA / hourly 2347 / notes.md §135 / Open-Jev densify HEAD f46ff604f794 / README SHA e32c4bbd519c / Publish audited JevBench public-subset baselines / public-subset ≠ Harbor / 231 ≠ 534 / jaredpalmer/kev densify HEAD e0bcf50153f1 / PLAN correct 35B MMLU-Pro (0.550) / evaluate.load honour weights_dtype=bf16 / wy-coliney/jev-browser-use 282★ / 5-10× *theirs* not Harbor / fail-open routing ≠ permission / ordered routing ≠ end-to-end / softmax next-token ≠ calibrated Noul / potential_match ≠ hiring decision / hourly 0049 / notes.md §136 / Open-Jev densify HEAD ed45657bf726 / README SHA 12e0f581e15d / Publish audited v3 community data and held-out evaluation protocol / Redesign readable project site and consolidate benchmark results / v3 data prepared ≠ retrained released models / held-out protocol ≠ Harbor / 1,280-row panel ≠ Harbor / finite training loss ≠ quality improvement / website redesign ≠ calibration / 129,288 decision rows 74,921 training / frozen mixture 96,849 training / 1,280-row / 840-group comparison panel / 27B step 616 pending / naive throws away 83% *theirs* / certo KL 0.008 acc 0.844 ECE 0.004 *theirs* / first-instinct 63.3%→78.1% *theirs* not Harbor / 371,278 prepared ≠ consumed / Jev is a gate not a generator / Lake remains admission / Jev never writes Lean / community port ≠ TypeSafe / chy4pro/jev-for-chrome ≠ browser-use/jev-ultrafast / joint RLCD *theirs* / Dohnuts ≠ TypeSafe / Akashdb5/jev-router ≠ gargpratyush/jev-router ≠ daviddl9/jev-router / kiuckhuang/laya-jev ≠ KonghaYao/laya-jev / tegersdorfer-collab/jevkit ≠ isiomaC/jevkit ≠ WaynezProg/jev-kit / buluoray/JevOnly already carded / yottayoshida/jev-intent-review already carded / skip-thin Iskandeur/system1-system2 zhlei07/open-system-one Hand-In/openjev-multimodal gwxcsny53/jev-watchtower empty SHA / hourly 0151 / notes.md §137 / GLiClass knowledgator Hub family class-peer catalog not Jev equivalent / typed-decision-leaderboard *theirs* not Harbor / JEV 0.7350 ZTC 27B 0.7289 / Jevbridge ACP and MCP adapter / Cut the slop / Not a Cua binding / open reproductions of the shape / 82.3% ECE 0.017 *theirs* / hourly 0248 / notes.md §138, or "cascade sign-flip / calibration theater": read `references/faq.md`, / Greedy 0.90 vs Oracle 0.82 *theirs* / Random conf 0.00 still 20.5% *theirs* / 8,400 calls $0.39 *theirs* / Noul 0.7 true 44% *theirs* / JevBench 81.65 *theirs* not Harbor / WindTunnel 49/49 *theirs* not Harbor / 0-byte Mandelbrot is not a replica / training not complete / Compose meaning like state / Code enumerates the candidates / Jev is the first classifier the design is bound to none / context is the conversation so far / A clean report is not proof / does not sandbox / ChatJEVs ≠ erik-dunteman/ChatJev / generation from Choice is not a language model replica / demo scores are not accuracy measurements / Qwen2.5 ≠ Archer / confidence ≠ P(correct) / seed 42 n=1 is not Harbor / meijustory123/OpenJev-Kit IS meijustory123/openjev / skip-thin Fibonaccirabbit/Jev-GalGame MadhavBahl/jev-guide advance-lion/dsh-jev-hooks amithgc/local-jev hiro1202/jev-review-gate-poc inlight37-design/decision-model_lab kuhung/ask-jev mmiguez314/jev-lab pomodorozhong/exp-jev vanthiet1/JevGuarAgent empty SHA / hourly 0348 / notes.md §139 / keeps essentially every block 0.0%/−0.5% *theirs* / Token reduction alone is not cost reduction / Chat 1282.3 ms vs gateway 232.1 ms ~1/5.5 *theirs* / acc 0.796 ECE 0.027 *theirs* / How you ask mattered more / Calibration is not yet measured / Add JevHarness project link to READMEs / densify §115 not a sibling first sighting / Awesomejev 691→802 / tracker likes 81 lastModified UNCHANGED / hf:AXERA-TECH/Laya / hf:openjev/openjev-MLX-4bit ≠ razorback16/openjev / GeekyAbs/laya ≠ convaiinnovations/laya / Softmax over candidate logprobs / skip-thin Adrian-lzr/jev-spire-brain Dililianxice/jev-robotic-arm-benchmark baltzparra/jev-study lzero07/jev-laya-statement qq150078158-lab/TDM-demo empty SHA / hourly 0445 / notes.md §140 then `references/mental-models.md`, then `references/mixed-architecture.md`, then `references/judgment-class.md` before any mapping. Proof, @@ -610,6 +610,21 @@ Hourly 0049 uniqueness lock: Zefan-Cai/Open-Jev densify HEAD f46ff604f794 via af **Hourly 0151 HIGH (`notes.md` §137).** Open-Jev v3 densify HEAD ed45657bf726. README SHA 12e0f581e15d. v3 data prepared ≠ retrained released models. held-out protocol ≠ Harbor. 1,280-row panel ≠ Harbor. finite training loss ≠ quality improvement. website redesign ≠ calibration. jev-wide naive throws away 83% *theirs*. certo KL 0.008 *theirs*. first-instinct 63.3%→78.1% *theirs* not Harbor. Jev is a gate not a generator. community port ≠ TypeSafe. catalog ≠ endorsement. *theirs* not Harbor. SHA move is not a replica. Do not reopen or amend PR #23–#60. Does not bump 0.5.0. Skip Archer. `invented_signal: false`. Hourly 0151 uniqueness lock: Zefan-Cai/Open-Jev densify HEAD ed45657bf726 via 748ae3024294 README SHA 12e0f581e15d was e32c4bbd519c; Publish audited v3 community data and held-out evaluation protocol; Redesign readable project site and consolidate benchmark results; 129,288 decision rows 74,921 training; frozen mixture 96,849 training; 1,280-row / 840-group comparison panel; v3 data prepared ≠ retrained released models; held-out protocol ≠ Harbor; 1,280-row panel ≠ Harbor; finite training loss ≠ quality improvement; website redesign ≠ calibration; 27B step 616 pending; Open-Jev TREC pending; LoRA ≠ RLCD replica; Qwen3.5-2B ≠ Archer; Qwen3.5-9B ≠ Archer; densify §125 not a sibling first sighting; chy4pro/jev-for-chrome 12★ community port ≠ TypeSafe; chy4pro/jev-for-chrome ≠ browser-use/jev-ultrafast; PsiACE/dohnuts 4★ small multimodal direct decisions; joint RLCD *theirs*; Dohnuts ≠ TypeSafe; catoenm/first-instinct 9B 63.3%→78.1% *theirs* not Harbor; 371,278 prepared ≠ consumed; RL did not reliably improve held-out; independent educational not a recovered Jev recipe; 123Satyajeet123/jev-wide naive throws away 83% *theirs*; 255 documented ~32,768 tokens real; two-decimal 95.8% floored *theirs*; IIA fails +0.31 ... +0.50 *theirs*; AltSlate-Labs/certo KL 0.008 acc 0.844 ECE 0.004 *theirs*; research preview independent not affiliated; endomorphosis/JevOps Jev is a gate not a generator; Lake remains admission; Jev never writes Lean; gbesse/question-forge held-out before winner; demo accuracy is synthetic not a Jev benchmark; flyryan/ai-news-aggregator 26★ does not execute; Akashdb5/jev-router ≠ gargpratyush/jev-router ≠ daviddl9/jev-router; kiuckhuang/laya-jev ≠ KonghaYao/laya-jev; tegersdorfer-collab/jevkit ≠ isiomaC/jevkit ≠ WaynezProg/jev-kit; buluoray/JevOnly already carded; yottayoshida/jev-intent-review already carded; skip-thin Iskandeur/system1-system2 zhlei07/open-system-one Hand-In/openjev-multimodal gwxcsny53/jev-watchtower empty SHA; serving substrate ≠ calibrated replica; game success ≠ calibrated Noul; does not execute; catalog ≠ endorsement; routing ≠ permission; *theirs* not Harbor; 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/#53/#54/#55/#56/#57/#58/#59/#60; notes.md §137 +## Hourly 0445 HIGH (`notes.md` §140) + +lcc keep-all / Token reduction alone is not cost reduction. +keeps essentially every block 0.0%/−0.5% *theirs*. +Chat 1282.3 ms vs gateway 232.1 ms ~1/5.5 *theirs*. +acc 0.796 ECE 0.027 *theirs*. How you ask mattered more. +Calibration is not yet measured. NanoJev densify JevHarness §115. +Awesomejev 691→802. serving substrate ≠ calibrated replica. +catalog ≠ endorsement. *theirs* not Harbor. +Do not copy keys. Fresh PR off `777546f` (merged #63). +Do not reopen or amend PR #23–#63. Does not bump 0.5.0. Skip Archer. +`invented_signal: false`. + +Hourly 0445 uniqueness lock: lucasmartins-ai/lcc 7★ MIT HEAD a7e86fb60997 README SHA 877831764be9; keeps essentially every block 0.0%/−0.5% *theirs*; mechanical −70.0% Jev −52.1% *theirs*; mock Laya = Jev −22.6% on XL withdrawn; Token reduction alone is not cost reduction; N=18 pilot not Harbor; David-Lolly/Jev-Compatible 3★ HEAD e52e963d8539 README SHA 6e5ff22d40a6; Chat 1282.3 ms vs gateway 232.1 ms ~1/5.5 *theirs*; 3/3 n=3; Softmax over candidate logprobs; Qwen3.8-27B ≠ Archer; wire-compat ≠ logit-equiv; hwfengcs/any2jev 2★ Apache-2.0 HEAD 719b0eb9eefe README SHA 27e0af212cd5; independent not affiliated; acc 0.796 ECE 0.027 *theirs*; 42 ms vs JSON 778 ms *theirs*; Snake acc 0.953 ECE 0.034 *theirs*; Qwen3-0.6B ≠ Archer; /v1/systemone wire-compat ≠ logit-equiv; TianyuCodings/NanoJev densify HEAD 76fdfc9ecdca README SHA a8f8afeb7e44 was 618cea6d / 4190093c64ee; Add JevHarness project link to READMEs; densify §115 not a sibling first sighting; SHA move is not a replica; jjd-lab/jev-synthetic-survey MIT HEAD 9ca8c4ab94bb README SHA ec1664288d50; How you ask mattered more; Noul TVD 0.1530 vs GPT 0.1789 *theirs*; ECE 0.1472 *theirs* not Harbor; missed 0.05 bar; 67.28% vs 64.78% *theirs*; $4.02 vs ~$136 *theirs*; independent work; CankatSarac/jev-arcade MIT HEAD b2e45ed3c1c6 README SHA 1f06af0f4c74; snake 70/80 *theirs*; tetris 167 vs heuristic 2333 *theirs*; 74% conf <0.5 *theirs*; Calibration is not yet measured; three seeds not Harbor; game success ≠ calibrated Noul; sszxt/rlcd HEAD 66ca01664d6b README SHA 3da08d46758d; ECE 0.490→0.423 Brier 0.487→0.409 *theirs*; Yang 2023 contrastive ≠ TypeSafe RLCD; Qwen2.5 ≠ Archer; still overconfident; hf:AXERA-TECH/Laya sha 51a586cd14e2 apache; AX650 NPU 69.991/27.722/69.990 ms *theirs*; seq 256 up to 4 options; serving substrate ≠ calibrated replica; base convaiinnovations/laya; hf:openjev/openjev-MLX-4bit sha c59bf1eed7d8 cc-by-nc-4.0; ~15 GB 4-bit affine; independent not affiliated; hf:openjev/openjev-MLX-4bit ≠ razorback16/openjev; hf:GeekyAbs/laya sha b65d05b4d9eb; GeekyAbs/laya ≠ convaiinnovations/laya; hf:alfred361/laya-web sha 33f171161da5; 100% argmax *theirs*; multilingual-int8 93.8% / worst shift 16.9 pts *theirs*; 50bbx/laya-needle Apache HEAD 01961bade52f README SHA ecaff4dfd271; threshold 0.58 still soft; local Laya ≠ hosted Jev; yunhai-dev/laya2typesafeapi HEAD 4aeb89be286b README SHA d2e5d114fcd8; TypeSafe-compatible ≠ TypeSafe replica; iamdgarcia/openJev MIT HEAD 62bbc30eece2 README SHA 55614ad8caab; independent educational; not local inference; iamdgarcia/openJev ≠ alongL/openJev ≠ Zefan-Cai/Open-Jev; chrisns/homebrew-laya-mac-serve MIT HEAD 1b3c4c0bdb70 README SHA 5e58082b7e9b; tap for chrisns/laya-mac-serve §139; serving substrate ≠ calibrated replica; nk412/judgements MIT HEAD 6624e53c86cc README SHA 86fe465f7887; pydantic wrapper; threshold 0.5 still soft; JingHao-Leon/awesome-jev-apps MIT HEAD d3ef0254c4b2 README SHA fd38a3c4ff2e; catalog ≠ endorsement; JingHao-Leon/awesome-jev-apps ≠ heyjunpenn/awesome-jev; Manta-Boardgame/jev-chat HEAD 44721bae8c2e README SHA 03272e4b9a9f; unofficial; 98% confidence *theirs*; ximing/jev-snake-game HEAD 1e80283f458e README SHA 0a54b74eb095; 用 TypeSafe Jev 驱动的自动贪吃蛇; hf:dataset:syvai/danish-dynaword-laya gated HTTP 401; size_categories 10K 1e-9 or abs(gateway_ms - 232.1) > 1e-9: + raise ValueError("unexpected latency") + return logit_equiv is False + + +def any2jev_ece_is_not_harbor(acc, ece, harbor=False): + """acc 0.796 ECE 0.027 *theirs*.""" + if abs(acc - 0.796) > 1e-9 or abs(ece - 0.027) > 1e-9: + raise ValueError("unexpected scores") + return harbor is False + + +def how_you_ask_mattered_more(kind, model_first=False): + """How you ask mattered more.""" + if kind != "synthetic_survey_ask": + raise ValueError("unexpected kind") + return model_first is False + + +def arcade_calibration_not_yet_measured(kind, measured=False): + """Calibration is not yet measured. three seeds not Harbor.""" + if kind != "jev_arcade": + raise ValueError("unexpected kind") + return measured is False + def clean_report_is_not_sandbox(kind, sandbox=False): """A clean report is not proof. does not sandbox.""" if kind != "is_malicious_scan": @@ -1334,6 +1370,24 @@ def self_test(): assert theirs_bench_is_not_harbor(39, "typesafe-ai-test-0.39") assert theirs_bench_is_not_harbor(4949, "werr-windtunnel-49-49") + # 0445: lcc keep-all is not cost reduction / softmax gateway is not + # logit-equiv / acc 0.796 ECE 0.027 *theirs* / How you ask mattered + # more / Calibration is not yet measured. + assert lcc_keep_all_is_not_cost_reduction("lcc_keep_all", False) + assert not lcc_keep_all_is_not_cost_reduction("lcc_keep_all", True) + assert softmax_gateway_is_not_logit_equiv(1282.3, 232.1, False) + assert not softmax_gateway_is_not_logit_equiv(1282.3, 232.1, True) + assert any2jev_ece_is_not_harbor(0.796, 0.027, False) + assert not any2jev_ece_is_not_harbor(0.796, 0.027, True) + assert how_you_ask_mattered_more("synthetic_survey_ask", False) + assert not how_you_ask_mattered_more("synthetic_survey_ask", True) + assert arcade_calibration_not_yet_measured("jev_arcade", False) + assert not arcade_calibration_not_yet_measured("jev_arcade", True) + assert theirs_bench_is_not_harbor(796, "any2jev-acc-0.796") + assert theirs_bench_is_not_harbor(2321, "jev-compatible-232.1ms") + assert theirs_bench_is_not_harbor(7080, "jev-arcade-snake-70-80") + + diff --git a/.agents/skills/augustus/scripts/uniqueness_gate.py b/.agents/skills/augustus/scripts/uniqueness_gate.py index 6ce3e6e..b9e458e 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), Open-Jev densify (§125), hourly 1946 HIGH (§131), hourly 2049 HIGH (§132), hourly 2146 HIGH (§133), hourly 2246 HIGH (§134), hourly 2347 HIGH (§135), hourly 0049 HIGH (§136), hourly 0151 HIGH (§137), hourly 0248 HIGH (§138), and hourly 0348 HIGH (§139). +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), hourly 1946 HIGH (§131), hourly 2049 HIGH (§132), hourly 2146 HIGH (§133), hourly 2246 HIGH (§134), hourly 2347 HIGH (§135), hourly 0049 HIGH (§136), hourly 0151 HIGH (§137), hourly 0248 HIGH (§138), hourly 0348 HIGH (§139), and hourly 0445 HIGH (§140). Each lock must appear as one consecutive substring in every listed overlay. Fragments scattered across files do not count. @@ -11,9 +11,9 @@ 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–§139; -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, 505–520, 521–536, 537–552, 553–568, 569–584, 585–600, 601–616, 617–632, and 633–648 exist; -findings batches #97–#121 exist. Items 317–321 stay unused. +Also: YAML-parse SKILL.md frontmatter; notes.md owns §114–§140; +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, 505–520, 521–536, 537–552, 553–568, 569–584, 585–600, 601–616, 617–632, 633–648, and 649–664 exist; +findings batches #97–#122 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). @@ -25,6 +25,7 @@ The 0151 archive run_digest must claim §137 / 601–616 / #119. The 0248 archive run_digest must claim §138 / 617–632 / #120. The 0348 archive run_digest must claim §139 / 633–648 / #121. +The 0445 archive run_digest must claim §140 / 649–664 / #122. CHANGELOG.md must not hold uniqueness dump walls (dumps live in changelog-hourly.md). README.md must not hold the 0743 dump wall. Pages greps stay in docs/index.md and docs/_layouts/default.html. @@ -219,6 +220,10 @@ "Hourly 0348 uniqueness lock: JonathanHHenson/open-cricket MIT HEAD d75af22125ed README SHA 7d288a741089; Local structured decisions using causal language models; default Qwen/Qwen2.5-1.5B-Instruct; independent of TypeSafe; API follows Jev's general call shapes but model predictions and confidence calibration differ; wire-compat ≠ logit-equiv; Qwen2.5 ≠ Archer; replica ≠ TypeSafe; virtualman333/jev-decision-arena MIT HEAD cf6ae4ed31e8 README SHA 037f75d9610d; Greedy 0.90 vs Oracle 0.82 *theirs*; Random conf 0.00 still 20.5% *theirs*; ECE 0.180 / 0.106 / 0.205 *theirs*; confidence ≠ P(correct); game success ≠ calibrated Noul; seed 42 n=1 is not Harbor; dopeCape/typesafe-ai-test HEAD ed2adb7740d7 README SHA 183f91c36471; 8,400 calls $0.39 *theirs*; Noul 0.7 true 44% *theirs*; ≥0.9 conf 91.7% AG News *theirs*; versioned model ids rejected; *theirs* not Harbor; pCwOrM/werr 2★ MIT HEAD 2526cae98891 README SHA b29476734a09; JevBench 81.65 *theirs* not Harbor; WindTunnel 49/49 *theirs* not Harbor; 0-byte Mandelbrot is not a replica; meijustory123/OpenJev-Kit HEAD c53125982f80 README SHA a4e72c61a973; training not complete; no accuracy; Qwen3.5-0.8B ≠ Archer; meijustory123/OpenJev-Kit IS meijustory123/openjev (same GitHub id 1379187719); meijustory123/OpenJev-Kit ≠ Zefan-Cai/Open-Jev; microchipgnu/jev-hooks HEAD cbf40e64d7b2 README SHA af25fb0aaf70; Compose meaning like state; rashedInt32/jury.nvim 1★ MIT HEAD bf31e9509e7e README SHA a2b088dd0787; Code enumerates the candidates; evoke-build/evoke 1★ Apache-2.0 HEAD 310840b56f1d README SHA 02b91962cef4; Jev is the first classifier the design is bound to none; moritzkremb/jev-voice-browser densify HEAD 198a0764395a README SHA 816309fc22e6 was fa033303; context is the conversation so far; densify §82 not a sibling first sighting; luantak/is-malicious densify 18★ MIT HEAD faf6ba61d7e1 README SHA 4ae098b4b7ae; A clean report is not proof; does not sandbox; skillseedorg/ChatJEVs MIT HEAD 346e7347cf90 README SHA 6ff81d54040f; ChatJEVs ≠ erik-dunteman/ChatJev; generation from Choice is not a language model replica; chrisns/laya-mac-serve MIT HEAD f294500821b6 README SHA 00e39a7d04e2; serving substrate ≠ calibrated replica; rimusz/localjev-mlx HEAD 297836a0d95e README SHA 2d96d20e0b80; rimusz/localjev-mlx ≠ githubnext/localjev; luhayes/jev-agent-router 1★ MIT HEAD bba795a4dc4e README SHA 17a2f44993d1; does not execute; cutoff 0.8 still soft; gbesse/decision-workbench MIT HEAD 8889cf3750a3 README SHA 14a3bf79da7a; demo scores are not accuracy measurements; zhuyansen/x-reply-filter already carded; kylemclaren/jev-search ≠ kazuhideoki/jev-search; xinwang-nwpu/jev-mobile ≠ Friedjof/jev-mobile; kcd-dev/jev-skill ≠ raphael-liu/jev-skill; yanmad27/ask-jev ≠ kuhung/ask-jev; hf:s1lv3rj1nx/openjev-router-healthcare encoder class member not Jev replica; hf:akhilaaa3/openjev-v1-40705-nimble-r512-merged ≠ hf:akhilaaa3/openjev-v1-allmix-r512-merged; jevai spaces catalog ≠ endorsement; skip-thin Fibonaccirabbit/Jev-GalGame MadhavBahl/jev-guide advance-lion/dsh-jev-hooks amithgc/local-jev hiro1202/jev-review-gate-poc inlight37-design/decision-model_lab kuhung/ask-jev mmiguez314/jev-lab pomodorozhong/exp-jev vanthiet1/JevGuarAgent empty SHA; catalog ≠ endorsement; game success ≠ calibrated Noul; does not execute; routing ≠ permission; *theirs* not Harbor; 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/#53/#54/#55/#56/#57/#58/#59/#60/#61/#62; notes.md §139" ) +UNIQ_0445 = ( + 'Hourly 0445 uniqueness lock: lucasmartins-ai/lcc 7★ MIT HEAD a7e86fb60997 README SHA 877831764be9; keeps essentially every block 0.0%/−0.5% *theirs*; mechanical −70.0% Jev −52.1% *theirs*; mock Laya = Jev −22.6% on XL withdrawn; Token reduction alone is not cost reduction; N=18 pilot not Harbor; David-Lolly/Jev-Compatible 3★ HEAD e52e963d8539 README SHA 6e5ff22d40a6; Chat 1282.3 ms vs gateway 232.1 ms ~1/5.5 *theirs*; 3/3 n=3; Softmax over candidate logprobs; Qwen3.8-27B ≠ Archer; wire-compat ≠ logit-equiv; hwfengcs/any2jev 2★ Apache-2.0 HEAD 719b0eb9eefe README SHA 27e0af212cd5; independent not affiliated; acc 0.796 ECE 0.027 *theirs*; 42 ms vs JSON 778 ms *theirs*; Snake acc 0.953 ECE 0.034 *theirs*; Qwen3-0.6B ≠ Archer; /v1/systemone wire-compat ≠ logit-equiv; TianyuCodings/NanoJev densify HEAD 76fdfc9ecdca README SHA a8f8afeb7e44 was 618cea6d / 4190093c64ee; Add JevHarness project link to READMEs; densify §115 not a sibling first sighting; SHA move is not a replica; jjd-lab/jev-synthetic-survey MIT HEAD 9ca8c4ab94bb README SHA ec1664288d50; How you ask mattered more; Noul TVD 0.1530 vs GPT 0.1789 *theirs*; ECE 0.1472 *theirs* not Harbor; missed 0.05 bar; 67.28% vs 64.78% *theirs*; $4.02 vs ~$136 *theirs*; independent work; CankatSarac/jev-arcade MIT HEAD b2e45ed3c1c6 README SHA 1f06af0f4c74; snake 70/80 *theirs*; tetris 167 vs heuristic 2333 *theirs*; 74% conf <0.5 *theirs*; Calibration is not yet measured; three seeds not Harbor; game success ≠ calibrated Noul; sszxt/rlcd HEAD 66ca01664d6b README SHA 3da08d46758d; ECE 0.490→0.423 Brier 0.487→0.409 *theirs*; Yang 2023 contrastive ≠ TypeSafe RLCD; Qwen2.5 ≠ Archer; still overconfident; hf:AXERA-TECH/Laya sha 51a586cd14e2 apache; AX650 NPU 69.991/27.722/69.990 ms *theirs*; seq 256 up to 4 options; serving substrate ≠ calibrated replica; base convaiinnovations/laya; hf:openjev/openjev-MLX-4bit sha c59bf1eed7d8 cc-by-nc-4.0; ~15 GB 4-bit affine; independent not affiliated; hf:openjev/openjev-MLX-4bit ≠ razorback16/openjev; hf:GeekyAbs/laya sha b65d05b4d9eb; GeekyAbs/laya ≠ convaiinnovations/laya; hf:alfred361/laya-web sha 33f171161da5; 100% argmax *theirs*; multilingual-int8 93.8% / worst shift 16.9 pts *theirs*; 50bbx/laya-needle Apache HEAD 01961bade52f README SHA ecaff4dfd271; threshold 0.58 still soft; local Laya ≠ hosted Jev; yunhai-dev/laya2typesafeapi HEAD 4aeb89be286b README SHA d2e5d114fcd8; TypeSafe-compatible ≠ TypeSafe replica; iamdgarcia/openJev MIT HEAD 62bbc30eece2 README SHA 55614ad8caab; independent educational; not local inference; iamdgarcia/openJev ≠ alongL/openJev ≠ Zefan-Cai/Open-Jev; chrisns/homebrew-laya-mac-serve MIT HEAD 1b3c4c0bdb70 README SHA 5e58082b7e9b; tap for chrisns/laya-mac-serve §139; serving substrate ≠ calibrated replica; nk412/judgements MIT HEAD 6624e53c86cc README SHA 86fe465f7887; pydantic wrapper; threshold 0.5 still soft; JingHao-Leon/awesome-jev-apps MIT HEAD d3ef0254c4b2 README SHA fd38a3c4ff2e; catalog ≠ endorsement; JingHao-Leon/awesome-jev-apps ≠ heyjunpenn/awesome-jev; Manta-Boardgame/jev-chat HEAD 44721bae8c2e README SHA 03272e4b9a9f; unofficial; 98% confidence *theirs*; ximing/jev-snake-game HEAD 1e80283f458e README SHA 0a54b74eb095; 用 TypeSafe Jev 驱动的自动贪吃蛇; hf:dataset:syvai/danish-dynaword-laya gated HTTP 401; size_categories 10K int: failed.append(f"0248 lock missing as one substring: {rel}") if UNIQ_0348 not in body: failed.append(f"0348 lock missing as one substring: {rel}") + if UNIQ_0445 not in body: + failed.append(f"0445 lock missing as one substring: {rel}") if "meijustory123/OpenJev-Kit ≠ meijustory123/openjev" in body: failed.append( "0348 false namesake lock still present " @@ -342,6 +349,10 @@ def main() -> int: ) if "meijustory123/OpenJev-Kit IS meijustory123/openjev (same GitHub id 1379187719)" not in UNIQ_0348: failed.append("UNIQ_0348 missing OpenJev-Kit IS openjev same-id lock") + if "GeekyAbs/laya ≠ convaiinnovations/laya" not in UNIQ_0445: + failed.append("UNIQ_0445 missing GeekyAbs/laya ≠ convaiinnovations/laya") + if "densify §115 not a sibling first sighting" not in UNIQ_0445: + failed.append("UNIQ_0445 missing NanoJev densify §115 lock") for rel in REVISIT_OVERLAYS: path = ROOT / rel if not path.is_file(): @@ -403,10 +414,12 @@ def main() -> int: failed.append("notes.md missing §138 heading") if "## 139. Hourly 0348 HIGH" not in notes: failed.append("notes.md missing §139 heading") + if "## 140. Hourly 0445 HIGH" not in notes: + failed.append("notes.md missing §140 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)) + list(range(505, 521)) + list(range(521, 537)) + list(range(537, 553)) + list(range(553, 569)) + list(range(569, 585)) + list(range(585, 601)) + list(range(601, 617)) + list(range(617, 633)) + list(range(633, 649)): + 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)) + list(range(521, 537)) + list(range(537, 553)) + list(range(553, 569)) + list(range(569, 585)) + list(range(585, 601)) + list(range(601, 617)) + list(range(617, 633)) + list(range(633, 649)) + list(range(649, 665)): needle = f"{n}. **" if needle not in algebra: failed.append(f"composition-algebra missing item {n}") @@ -441,6 +454,7 @@ def main() -> int: "## Batch #119", "## Batch #120", "## Batch #121", + "## Batch #122", ): if batch not in findings: failed.append(f"findings.md missing {batch}") @@ -549,6 +563,27 @@ def main() -> int: ) if digest0348.get("invented_signal") is not False: failed.append("0348 run_digest invented_signal is not false") + digest_path_0445 = ROOT / "research/archive/hourly/2026-09-21T11/run_digest.json" + if not digest_path_0445.is_file(): + failed.append("missing 0445 run_digest.json") + else: + digest0445 = json.loads(digest_path_0445.read_text(encoding="utf-8")) + if digest0445.get("label") != "0445": + failed.append(f"0445 run_digest label {digest0445.get('label')!r} != '0445'") + if digest0445.get("notes_section") != "140": + failed.append( + f"0445 run_digest notes_section {digest0445.get('notes_section')!r} != '140'" + ) + if digest0445.get("composition") != "649-664": + failed.append( + f"0445 run_digest composition {digest0445.get('composition')!r} != '649-664'" + ) + if digest0445.get("findings_batch") != 122: + failed.append( + f"0445 run_digest findings_batch {digest0445.get('findings_batch')!r} != 122" + ) + if digest0445.get("invented_signal") is not False: + failed.append("0445 run_digest invented_signal is not false") digest_path_0151 = ROOT / "research/archive/hourly/2026-09-21T08/run_digest.json" if not digest_path_0151.is_file(): failed.append("missing 0151 run_digest.json") @@ -1107,6 +1142,22 @@ def main() -> int: 'seed 42 n=1 is not Harbor', 'skip-thin Fibonaccirabbit/Jev-GalGame MadhavBahl/jev-guide advance-lion/dsh-jev-hooks amithgc/local-jev hiro1202/jev-review-gate-poc inlight37-design/decision-model_lab kuhung/ask-jev mmiguez314/jev-lab pomodorozhong/exp-jev vanthiet1/JevGuarAgent empty SHA', 'hourly 0348 / notes.md §139', + 'keeps essentially every block 0.0%/−0.5% *theirs*', + 'Token reduction alone is not cost reduction', + 'Chat 1282.3 ms vs gateway 232.1 ms ~1/5.5 *theirs*', + 'acc 0.796 ECE 0.027 *theirs*', + 'How you ask mattered more', + 'Calibration is not yet measured', + 'Add JevHarness project link to READMEs', + 'densify §115 not a sibling first sighting', + 'Awesomejev 691→802', + 'tracker likes 81 lastModified UNCHANGED', + 'hf:AXERA-TECH/Laya', + 'hf:openjev/openjev-MLX-4bit ≠ razorback16/openjev', + 'GeekyAbs/laya ≠ convaiinnovations/laya', + 'Softmax over candidate logprobs', + 'skip-thin Adrian-lzr/jev-spire-brain Dililianxice/jev-robotic-arm-benchmark baltzparra/jev-study lzero07/jev-laya-statement qq150078158-lab/TDM-demo empty SHA', + 'hourly 0445 / notes.md §140', ): if frag not in haystack: failed.append(f"SKILL.md missing fragment {frag!r}") @@ -1480,6 +1531,22 @@ def main() -> int: 'A clean report is not proof', 'does not sandbox', 'hourly 0348 / notes.md §139', + 'keeps essentially every block 0.0%/−0.5% *theirs*', + 'Token reduction alone is not cost reduction', + 'Chat 1282.3 ms vs gateway 232.1 ms ~1/5.5 *theirs*', + 'acc 0.796 ECE 0.027 *theirs*', + 'How you ask mattered more', + 'Calibration is not yet measured', + 'Add JevHarness project link to READMEs', + 'densify §115 not a sibling first sighting', + 'Awesomejev 691→802', + 'tracker likes 81 lastModified UNCHANGED', + 'hf:AXERA-TECH/Laya', + 'hf:openjev/openjev-MLX-4bit ≠ razorback16/openjev', + 'GeekyAbs/laya ≠ convaiinnovations/laya', + 'Softmax over candidate logprobs', + 'skip-thin Adrian-lzr/jev-spire-brain Dililianxice/jev-robotic-arm-benchmark baltzparra/jev-study lzero07/jev-laya-statement qq150078158-lab/TDM-demo empty SHA', + 'hourly 0445 / notes.md §140', ): if frag not in proto_line: failed.append(f"SKILL.md protocol missing {frag!r}") @@ -1511,6 +1578,7 @@ def main() -> int: ("0151", UNIQ_0151), ("0248", UNIQ_0248), ("0348", UNIQ_0348), + ("0445", UNIQ_0445), ): if lock in changelog: failed.append( @@ -1594,7 +1662,7 @@ def main() -> int: f"2146 chars={len(UNIQ_2146)} " f"2246 chars={len(UNIQ_2246)} " f"2347 chars={len(UNIQ_2347)} " - f"0049 chars={len(UNIQ_0049)} 0151 chars={len(UNIQ_0151)} 0248 chars={len(UNIQ_0248)} 0348 chars={len(UNIQ_0348)} " + f"0049 chars={len(UNIQ_0049)} 0151 chars={len(UNIQ_0151)} 0248 chars={len(UNIQ_0248)} 0348 chars={len(UNIQ_0348)} 0445 chars={len(UNIQ_0445)} " 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 6939fc5..b212abf 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -16,6 +16,43 @@ folds: `research/notes.md`. ## [Unreleased] +Hourly 0445 HIGH (`research/notes.md` §140 / composition items +649–664 / findings batch #122). 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–#63. +Do not amend released 0.5.0 (#42). Merged #63 owns §139. Merged #62 +owns §138. Merged #61 owns §137. + +### Added + +- **Hourly 0445 HIGH (`notes.md` §140).** lcc real Laya keep-all. + Token reduction alone is not cost reduction. keeps essentially every + block 0.0%/−0.5% *theirs*. mock Laya = Jev −22.6% on XL withdrawn. + Chat 1282.3 ms vs gateway 232.1 ms ~1/5.5 *theirs*. Softmax over + candidate logprobs. any2jev acc 0.796 ECE 0.027 *theirs*. How you + ask mattered more. Calibration is not yet measured. NanoJev densify + JevHarness §115. Awesomejev 691→802 quote watch not re-derive. + serving substrate ≠ calibrated replica. skip-thin empty SHA. + Evaluator: lcc keep-all is not cost reduction / softmax gateway is + not logit-equiv / acc 0.796 ECE 0.027 *theirs* / How you ask mattered + more / Calibration is not yet measured. + 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 + 2049 + 2146 + 2246 + 2347 + 0049 + 0151 + + 0248 + 0348 + 0445. + Composition items 649–664 / batch #122. + **HARD RULE:** do not reopen or amend PR #23–#63. Does **not** bump + 0.5.0. + +- **Recipe (class, not Jev-only).** Without Augustus: treat lcc keep-all + as cost reduction, 232.1 ms as logit-equiv, ECE 0.027 as Harbor, or + arcade 70/80 as calibrated Noul. With Augustus: Token reduction alone + is not cost reduction; wire-compat ≠ logit-equiv; Softmax over options + ≠ calibrated Noul; How you ask mattered more; Calibration is not yet + measured; densify §115 not a sibling first sighting; *theirs* not Harbor. + + Hourly 0348 HIGH (`research/notes.md` §139 / composition items 633–648 / findings batch #121). Does **not** bump the 0.5.0 pin. Uniqueness dumps live in diff --git a/README.md b/README.md index 6619389..acc8128 100644 --- a/README.md +++ b/README.md @@ -148,6 +148,7 @@ User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jc User-provided 0922 uniqueness lock: SemIf was formerly OpenJev; independent; not affiliated with Jev or TypeSafe; homepage openjev.com; default master; MIT; HEAD ca3ba65f1429; Tolerate float roundoff in MLX evidence verification; pushed 2026-09-19; live REST 2282★ / 140 forks; size 9177; README SHA 74ab7f7f; LICENSE SHA ca562883; interface pattern reproduction with open models; does not reproduce Jev undisclosed model/training; Direct option logits; 0 output tokens; shared-state parallel; MLX backend for Apple Silicon (`--backend mlx`); Qwen3.5-4B 3090 direct 1.023s vs AR JSON 5.332s (**5.21×**); argmax agree 18/21; systems comparison ≠ semantic equivalence; Parallel suffixes 20.03 dec/s on 777 decisions; Browser ladder Qwen3.5-4B authored BA 0.813, pert 0.766, TypeSafe subset agreement 0.845 vs Published Jev 0.883 (102 across 20 cases); Softmax over options ≠ calibrated Noul; typed output does not guarantee semantic correctness; wire/agreement ≠ replica of TypeSafe; SemIf ≠ kw2828/OpenJev playground ≠ zhihz/openjev ≠ apiplant/semif-rs port ≠ dddanielliu/semif-serve; rename is densify not a second census; JevBench 74.6 is §78 not this ladder; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#38; do not push onto open #39/#40; notes.md §117 User-provided 0940 uniqueness lock: Turn decision-shaped LLM prompts into proposed Jev primitives; This is a conversion assistant, not an automatic guarantee of equivalent behavior; The compiler uses deterministic heuristics, not an LLM or evaluation model; It understands a deliberately small set of common prompt patterns; Generated instructions and criteria must be reviewed before production use; Score ranges such as 0 to 1 are translated into ordered Jev criteria; Prompts requiring open-ended prose are not a fit; suitability strong/partial/not_a_fit; compatibility full/partial/none; Writing new text stays with an LLM; Review the generated Score rubric; Jev scores ordered criteria, not an arbitrary 0-to-1 range; Everything runs locally in the browser; There is no framework, database, account, API, or server-side prompt processing; The key is read from the process environment and is never stored or printed; connect-src 'none'; alexwestco/llm-to-jev ≠ altryne/jevify ≠ ryana/jevify ≠ fidecastro/jevify ≠ Mintzs/jevify ≠ gulagala001/jevify ≠ uspraveen/Jevify; HEAD 234058ab372d; README SHA 43cd94fb; LICENSE SHA 5f334006; compiler SHA fdf235d0; 2★; MIT; JavaScript; size 29; Pages https://alexwestco.github.io/llm-to-jev/; invented_signal false; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35; notes.md §118 Hourly 0947 uniqueness lock: Fast and cheap agent evals. jev as judge.; 18,041 skills from the 200 most-starred repos; Not a security scanner; 最简 Jev 调用演示器; confidence 不是正确率; q93304989-bit/jev-lab ≠ tanayvasishtha/jev-lab ≠ dairui1/jev-lab ≠ BrendanH18/jev-lab ≠ yibie/laya-jev-lab; 75% cheaper and 18% faster withdrawn; jev @0.15 100% recall 87% savings; 33Audits/jev-auto ≠ gargpratyush/jev-router; no Typesafe key, no PI_API_BASE, zero deps; tool-emitted Score/Noul ≠ calibrated Noul; semantic_compatibility: false; candidate_mass; Qwen3.5-2B ≠ Archer; Jev evaluates decisions; it cannot run a coding-agent session; Status: no model yet; S1LV3RJ1NX/openjev ≠ TheoLeeCJ/openjev; 28 accepted decisions; 3 targets; score 800; health 100; arcade game not a flight trainer; A successful live TypeSafe call has not been verified for v0.1.0; abhibansal60/tidy ≠ MANISH007700/tidy; No model, Jev included, predicted which channels its owner keeps; seed 1 selected on a held-out 400-item validation split; Brier 0.342 → 0.378; more accurate and more overconfident; Qwen3.5-4B ≠ Archer; static quants of kushalpatil/jevify-gemma4-26b-a4b; The labels were corrected, and one earlier result was retracted; zero of 23,869 eligible rows; Do not compare cost without checking task success; Exit 1 is not a proof; kisshan13/typesafe-ai-go ≠ Nibir1/typesafe-go ≠ official; 38 tests that cannot fail in a 356-model warehouse; if a parser can answer it, Jev is never asked; 359 of them; Games & Simulation 82; Education & Learning 1; Ratings are heuristics; syedabbasshaheer-art/jev-atlas ≠ ZeroX-01/jev-atlas ≠ Zaious/jev-capability-atlas ≠ gorock007/jev-atlas; anandi1989/awesome-jev-usecases ≠ whyashthakker/awesome-jev-use-cases ≠ walidboulanouar/awesome-jev-use-cases ≠ vamsikrishna2421/jev-usecases; Every headline result above is self-reported; Archer Hume 84.6% MMLU-Pro is a third-party probe not landed Archer; catalog ≠ endorsement; judge ≠ actuator; softmax over A–H ≠ Noul; SemIf 2270★; jevlike 1054★; TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 538★; Laya likes 889; tracker likes 68 lastModified UNCHANGED; Blackwood likes 2 gated manual; 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/#35/#36/#37/#38/#40; do not push onto open #39; notes.md §119 +- Hourly 0445 HIGH (`research/notes.md` §140 / items 649–664 / batch #122). lcc keep-all vs withdrawn mock / any2jev acc 0.796 ECE 0.027 *theirs* / Jev-Compatible 232.1 ms *theirs* / NanoJev JevHarness densify §115 / How you ask mattered more / Calibration is not yet measured. uniqueness_gate 0843+0915+jcr+0922+0940+0947+1049+1143+1248+1340+1441+1542+1643+1746+1843+1936+1946+2049+2146+2246+2347+0049+0151+0248+0348+0445. Does not bump 0.5.0. Merged #63 owns §139. Merged #62 owns §138. - Hourly 0348 HIGH (`research/notes.md` §139 / items 633–648 / batch #121). open-cricket BYOM wire-compat ≠ logit-equiv / Greedy 0.90 vs Oracle 0.82 *theirs* / confidence ≠ P(correct) / JevBench 81.65 *theirs* not Harbor / 0-byte Mandelbrot is not a replica. uniqueness_gate 0843+0915+jcr+0922+0940+0947+1049+1143+1248+1340+1441+1542+1643+1746+1843+1936+1946+2049+2146+2246+2347+0049+0151+0248+0348. Does not bump 0.5.0. Merged #62 owns §138. Merged #61 owns §137. - Hourly 0248 HIGH (`research/notes.md` §138 / items 617–632 / batch #120). GLiClass Hub class-peer not Jev equivalent / typed-decision-leaderboard *theirs* not Harbor / Jevbridge wire-compat ≠ logit-equiv / Not a Cua binding. uniqueness_gate 0843+0915+jcr+0922+0940+0947+1049+1143+1248+1340+1441+1542+1643+1746+1843+1936+1946+2049+2146+2246+2347+0049+0151+0248. Does not bump 0.5.0. Merged #61 owns §137. Merged #60 owns §136. - Hourly 1643 HIGH (`research/notes.md` §127 / items 449–464 / batch #109). openjev 0.3.0 densify / restructured vLLM head ≠ logit-equiv / typed judgments not opinions / Thresholds are policy not model. uniqueness_gate 0843+0915+jcr+0922+0940+0947+1049+1143+1248+1340+1441+1542+1643. Does not bump 0.5.0. Merged #49 owns §126. Merged #48 owns §125. @@ -209,3 +210,4 @@ Hourly 0049 uniqueness lock: Zefan-Cai/Open-Jev densify HEAD f46ff604f794 via af Hourly 0151 uniqueness lock: Zefan-Cai/Open-Jev densify HEAD ed45657bf726 via 748ae3024294 README SHA 12e0f581e15d was e32c4bbd519c; Publish audited v3 community data and held-out evaluation protocol; Redesign readable project site and consolidate benchmark results; 129,288 decision rows 74,921 training; frozen mixture 96,849 training; 1,280-row / 840-group comparison panel; v3 data prepared ≠ retrained released models; held-out protocol ≠ Harbor; 1,280-row panel ≠ Harbor; finite training loss ≠ quality improvement; website redesign ≠ calibration; 27B step 616 pending; Open-Jev TREC pending; LoRA ≠ RLCD replica; Qwen3.5-2B ≠ Archer; Qwen3.5-9B ≠ Archer; densify §125 not a sibling first sighting; chy4pro/jev-for-chrome 12★ community port ≠ TypeSafe; chy4pro/jev-for-chrome ≠ browser-use/jev-ultrafast; PsiACE/dohnuts 4★ small multimodal direct decisions; joint RLCD *theirs*; Dohnuts ≠ TypeSafe; catoenm/first-instinct 9B 63.3%→78.1% *theirs* not Harbor; 371,278 prepared ≠ consumed; RL did not reliably improve held-out; independent educational not a recovered Jev recipe; 123Satyajeet123/jev-wide naive throws away 83% *theirs*; 255 documented ~32,768 tokens real; two-decimal 95.8% floored *theirs*; IIA fails +0.31 ... +0.50 *theirs*; AltSlate-Labs/certo KL 0.008 acc 0.844 ECE 0.004 *theirs*; research preview independent not affiliated; endomorphosis/JevOps Jev is a gate not a generator; Lake remains admission; Jev never writes Lean; gbesse/question-forge held-out before winner; demo accuracy is synthetic not a Jev benchmark; flyryan/ai-news-aggregator 26★ does not execute; Akashdb5/jev-router ≠ gargpratyush/jev-router ≠ daviddl9/jev-router; kiuckhuang/laya-jev ≠ KonghaYao/laya-jev; tegersdorfer-collab/jevkit ≠ isiomaC/jevkit ≠ WaynezProg/jev-kit; buluoray/JevOnly already carded; yottayoshida/jev-intent-review already carded; skip-thin Iskandeur/system1-system2 zhlei07/open-system-one Hand-In/openjev-multimodal gwxcsny53/jev-watchtower empty SHA; serving substrate ≠ calibrated replica; game success ≠ calibrated Noul; does not execute; catalog ≠ endorsement; routing ≠ permission; *theirs* not Harbor; 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/#53/#54/#55/#56/#57/#58/#59/#60; notes.md §137 Hourly 0248 uniqueness lock: hf:knowledgator/gliclass-instruct-large-v1.0 43 likes sha 825e5478c1bf apache-2.0; Efficient zero-shot and few-shot multi-task model via sequence classification; GLiClass knowledgator Hub family class-peer catalog not Jev equivalent; Knowledgator/GLiClass.c already §123; Hub models first card as class-peer entries; GLiNER/GLiClass ports are class members not Jev replicas; hf:space:mayafree/typed-decision-leaderboard 33 likes sha f4fc44077818; typed-decision-leaderboard *theirs* not Harbor; JEV 0.7350 ZTC 27B 0.7289 ZTC 397B 0.7272 *theirs* not Harbor; 2,018 items same labels; three-way tie; tacticocc/Jevbridge 33★ MIT HEAD da443ea453ac README SHA 2178333c4c3b; Jevbridge ACP and MCP adapter; does not generate text; Any LLM as System One; wire-compat ≠ logit-equiv; tshmieldev/sharp 29★ MIT HEAD 17cbd8d9cc9e README SHA 783a5cde519c; Cut the slop; Filter your X timeline; kavehmz/typesafe-playground 11★ HEAD 733991a2924a README SHA 04c0b1f6e7da; real API calls not polished benchmarks; himomohi/aside-jev 7★ MIT HEAD e570db43b0e1 README SHA 288e7c91c307; Jev picks the next action from your defined candidates; Not a Cua binding; Jev is the model Aside is the browser runtime; nico-martin/open-jev 6★ MIT HEAD 52667199e8a5 README SHA 81c0485d5833; open reproductions of the shape; Nothing is generated; nico-martin/open-jev ≠ razorback16/openjev ≠ Zefan-Cai/Open-Jev ≠ meijustory123/openjev; hf:chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF 82.3% ECE 0.017 *theirs*; Same decision as bf16 94.4% *theirs*; Qwen3.5-2B ≠ Archer; serving substrate ≠ calibrated replica; hf:pngwn/nanodiff-350m-typed-decisions ECE 0.065 → 0.036 *theirs*; hf:litert-community/laya-LiteRT 144/144 *theirs*; gargpratyush/journey-evals A page that says Success is never accepted as proof; mpnikhil/dev-0.4b Banking77 91.33% BoolQ 85.20% *theirs*; encoder class member not Jev replica; n4ze3m/typed-decisions-synth 7,414 cases 25,859 questions; Nobody checked it; Zaious/jev-capability-atlas already carded; LocalLLaMA/typed-decisions already carded; fengyiqicoder/jevfeed already carded; Zhao-Tian-yi/awesome-jev ≠ Gerry9000/awesome-jev ≠ heyjunpenn/awesome-jev ≠ yibie/awesome-jev; kaustav1996/reflex ≠ vuckuola619/reflex; tphakala/jev-mcp ≠ jkudish/jev-mcp; ninthspace/hunch ≠ carldaws/hunch ≠ tpellet/hunch; ruban-24/switchboard ≠ cannacre8ive/switchboard-ai; hf:openjev/openjev ≠ razorback16/openjev; catalog ≠ endorsement; skip-thin Fibonaccirabbit/Jev-VLN imanshu03/jev-browser-use luca-saggese/laya.c empty SHA; game success ≠ calibrated Noul; does not execute; routing ≠ permission; *theirs* not Harbor; 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/#53/#54/#55/#56/#57/#58/#59/#60/#61; notes.md §138 Hourly 0348 uniqueness lock: JonathanHHenson/open-cricket MIT HEAD d75af22125ed README SHA 7d288a741089; Local structured decisions using causal language models; default Qwen/Qwen2.5-1.5B-Instruct; independent of TypeSafe; API follows Jev's general call shapes but model predictions and confidence calibration differ; wire-compat ≠ logit-equiv; Qwen2.5 ≠ Archer; replica ≠ TypeSafe; virtualman333/jev-decision-arena MIT HEAD cf6ae4ed31e8 README SHA 037f75d9610d; Greedy 0.90 vs Oracle 0.82 *theirs*; Random conf 0.00 still 20.5% *theirs*; ECE 0.180 / 0.106 / 0.205 *theirs*; confidence ≠ P(correct); game success ≠ calibrated Noul; seed 42 n=1 is not Harbor; dopeCape/typesafe-ai-test HEAD ed2adb7740d7 README SHA 183f91c36471; 8,400 calls $0.39 *theirs*; Noul 0.7 true 44% *theirs*; ≥0.9 conf 91.7% AG News *theirs*; versioned model ids rejected; *theirs* not Harbor; pCwOrM/werr 2★ MIT HEAD 2526cae98891 README SHA b29476734a09; JevBench 81.65 *theirs* not Harbor; WindTunnel 49/49 *theirs* not Harbor; 0-byte Mandelbrot is not a replica; meijustory123/OpenJev-Kit HEAD c53125982f80 README SHA a4e72c61a973; training not complete; no accuracy; Qwen3.5-0.8B ≠ Archer; meijustory123/OpenJev-Kit IS meijustory123/openjev (same GitHub id 1379187719); meijustory123/OpenJev-Kit ≠ Zefan-Cai/Open-Jev; microchipgnu/jev-hooks HEAD cbf40e64d7b2 README SHA af25fb0aaf70; Compose meaning like state; rashedInt32/jury.nvim 1★ MIT HEAD bf31e9509e7e README SHA a2b088dd0787; Code enumerates the candidates; evoke-build/evoke 1★ Apache-2.0 HEAD 310840b56f1d README SHA 02b91962cef4; Jev is the first classifier the design is bound to none; moritzkremb/jev-voice-browser densify HEAD 198a0764395a README SHA 816309fc22e6 was fa033303; context is the conversation so far; densify §82 not a sibling first sighting; luantak/is-malicious densify 18★ MIT HEAD faf6ba61d7e1 README SHA 4ae098b4b7ae; A clean report is not proof; does not sandbox; skillseedorg/ChatJEVs MIT HEAD 346e7347cf90 README SHA 6ff81d54040f; ChatJEVs ≠ erik-dunteman/ChatJev; generation from Choice is not a language model replica; chrisns/laya-mac-serve MIT HEAD f294500821b6 README SHA 00e39a7d04e2; serving substrate ≠ calibrated replica; rimusz/localjev-mlx HEAD 297836a0d95e README SHA 2d96d20e0b80; rimusz/localjev-mlx ≠ githubnext/localjev; luhayes/jev-agent-router 1★ MIT HEAD bba795a4dc4e README SHA 17a2f44993d1; does not execute; cutoff 0.8 still soft; gbesse/decision-workbench MIT HEAD 8889cf3750a3 README SHA 14a3bf79da7a; demo scores are not accuracy measurements; zhuyansen/x-reply-filter already carded; kylemclaren/jev-search ≠ kazuhideoki/jev-search; xinwang-nwpu/jev-mobile ≠ Friedjof/jev-mobile; kcd-dev/jev-skill ≠ raphael-liu/jev-skill; yanmad27/ask-jev ≠ kuhung/ask-jev; hf:s1lv3rj1nx/openjev-router-healthcare encoder class member not Jev replica; hf:akhilaaa3/openjev-v1-40705-nimble-r512-merged ≠ hf:akhilaaa3/openjev-v1-allmix-r512-merged; jevai spaces catalog ≠ endorsement; skip-thin Fibonaccirabbit/Jev-GalGame MadhavBahl/jev-guide advance-lion/dsh-jev-hooks amithgc/local-jev hiro1202/jev-review-gate-poc inlight37-design/decision-model_lab kuhung/ask-jev mmiguez314/jev-lab pomodorozhong/exp-jev vanthiet1/JevGuarAgent empty SHA; catalog ≠ endorsement; game success ≠ calibrated Noul; does not execute; routing ≠ permission; *theirs* not Harbor; 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/#53/#54/#55/#56/#57/#58/#59/#60/#61/#62; notes.md §139 +Hourly 0445 uniqueness lock: lucasmartins-ai/lcc 7★ MIT HEAD a7e86fb60997 README SHA 877831764be9; keeps essentially every block 0.0%/−0.5% *theirs*; mechanical −70.0% Jev −52.1% *theirs*; mock Laya = Jev −22.6% on XL withdrawn; Token reduction alone is not cost reduction; N=18 pilot not Harbor; David-Lolly/Jev-Compatible 3★ HEAD e52e963d8539 README SHA 6e5ff22d40a6; Chat 1282.3 ms vs gateway 232.1 ms ~1/5.5 *theirs*; 3/3 n=3; Softmax over candidate logprobs; Qwen3.8-27B ≠ Archer; wire-compat ≠ logit-equiv; hwfengcs/any2jev 2★ Apache-2.0 HEAD 719b0eb9eefe README SHA 27e0af212cd5; independent not affiliated; acc 0.796 ECE 0.027 *theirs*; 42 ms vs JSON 778 ms *theirs*; Snake acc 0.953 ECE 0.034 *theirs*; Qwen3-0.6B ≠ Archer; /v1/systemone wire-compat ≠ logit-equiv; TianyuCodings/NanoJev densify HEAD 76fdfc9ecdca README SHA a8f8afeb7e44 was 618cea6d / 4190093c64ee; Add JevHarness project link to READMEs; densify §115 not a sibling first sighting; SHA move is not a replica; jjd-lab/jev-synthetic-survey MIT HEAD 9ca8c4ab94bb README SHA ec1664288d50; How you ask mattered more; Noul TVD 0.1530 vs GPT 0.1789 *theirs*; ECE 0.1472 *theirs* not Harbor; missed 0.05 bar; 67.28% vs 64.78% *theirs*; $4.02 vs ~$136 *theirs*; independent work; CankatSarac/jev-arcade MIT HEAD b2e45ed3c1c6 README SHA 1f06af0f4c74; snake 70/80 *theirs*; tetris 167 vs heuristic 2333 *theirs*; 74% conf <0.5 *theirs*; Calibration is not yet measured; three seeds not Harbor; game success ≠ calibrated Noul; sszxt/rlcd HEAD 66ca01664d6b README SHA 3da08d46758d; ECE 0.490→0.423 Brier 0.487→0.409 *theirs*; Yang 2023 contrastive ≠ TypeSafe RLCD; Qwen2.5 ≠ Archer; still overconfident; hf:AXERA-TECH/Laya sha 51a586cd14e2 apache; AX650 NPU 69.991/27.722/69.990 ms *theirs*; seq 256 up to 4 options; serving substrate ≠ calibrated replica; base convaiinnovations/laya; hf:openjev/openjev-MLX-4bit sha c59bf1eed7d8 cc-by-nc-4.0; ~15 GB 4-bit affine; independent not affiliated; hf:openjev/openjev-MLX-4bit ≠ razorback16/openjev; hf:GeekyAbs/laya sha b65d05b4d9eb; GeekyAbs/laya ≠ convaiinnovations/laya; hf:alfred361/laya-web sha 33f171161da5; 100% argmax *theirs*; multilingual-int8 93.8% / worst shift 16.9 pts *theirs*; 50bbx/laya-needle Apache HEAD 01961bade52f README SHA ecaff4dfd271; threshold 0.58 still soft; local Laya ≠ hosted Jev; yunhai-dev/laya2typesafeapi HEAD 4aeb89be286b README SHA d2e5d114fcd8; TypeSafe-compatible ≠ TypeSafe replica; iamdgarcia/openJev MIT HEAD 62bbc30eece2 README SHA 55614ad8caab; independent educational; not local inference; iamdgarcia/openJev ≠ alongL/openJev ≠ Zefan-Cai/Open-Jev; chrisns/homebrew-laya-mac-serve MIT HEAD 1b3c4c0bdb70 README SHA 5e58082b7e9b; tap for chrisns/laya-mac-serve §139; serving substrate ≠ calibrated replica; nk412/judgements MIT HEAD 6624e53c86cc README SHA 86fe465f7887; pydantic wrapper; threshold 0.5 still soft; JingHao-Leon/awesome-jev-apps MIT HEAD d3ef0254c4b2 README SHA fd38a3c4ff2e; catalog ≠ endorsement; JingHao-Leon/awesome-jev-apps ≠ heyjunpenn/awesome-jev; Manta-Boardgame/jev-chat HEAD 44721bae8c2e README SHA 03272e4b9a9f; unofficial; 98% confidence *theirs*; ximing/jev-snake-game HEAD 1e80283f458e README SHA 0a54b74eb095; 用 TypeSafe Jev 驱动的自动贪吃蛇; hf:dataset:syvai/danish-dynaword-laya gated HTTP 401; size_categories 10K Push to talk, choose a reviewed state in the local Safari fixture, perform one registered fixture operation, and verify the exact postcondition.\n\nOnly synthetic, local, reversible fixture data is allowed. Notes and generic Mac control are deferred.\n\n## Safety Boundaries\n\n- Push-to-talk only. No always-on listening.\n- No arbitrary shell, AppleScript, coordinate clicking, generic typing, arbitrary URLs, clipboard transfer, screenshots, vision, external websites, or uncontrolled navigation.\n- The bounded computer-use executor is limited to the versioned Safari fixture's two fixed reversible controls. The TextEdit workspace expansion is deferred and does not dispatch input. The executor does not enable generic desktop control.\n- Experimental desktop mode is retained only as a visibly deferred UI state. It cannot submit Hermes tasks or change the native capability registry in this checkpoint.\n- Jev may select only an exact locally constructed capability ID.\n- No automatic sending, deleting, purchasing, publishing, sharing, account changes, or privacy-sensitive actions.\n- Stop is local, idempotent, and always available.\n- Unknown native effects become `outcome_unknown` and are never replayed automatically.\n- Credentials belong in macOS Keychain, never in source, fixtures, logs, screenshots, or chat.\n\n## Jev Access Gate\n\nThe live selector is implemented and conditionally approved for the current bounded private prototype. Daniel confirmed the applicable account, DOO MADE approval, direct-client posture, refill setting, and retention conditions. The app still requires explicit Mac-side enablement and a Keychain credential before any provider request can occur; broader data and production use remain outside this approval.\n\nCredentials must be entered through the SecureField in **Jev Command Panel \u2192 Settings\u2026 \u2192 TypeSafe credential** and stored only in the Mac Keychain. Do not paste passwords, API keys, or tokens into chat or commit them. See `Docs/LIVE_JEV.md` and `Docs/DATA_EGRESS.md`.\n\n## Repository Map\n\n- `Design/`: UI/UX brief and acceptance checklist.\n- `Docs/`: product contract, capability registry, policy, egress, probes, tests, experiments, and publication clearance.\n- `Sources/JevCore/`: pure capability, policy, state, transcript, selection, and authority logic.\n- `Sources/JevMacShell/`: SwiftUI menu-bar shell, vis", + "readme_len": 5261, + "description_hash": "d24f745e25cd" + }, + { + "id": "David-Lolly/Jev-Compatible", + "source": "github", + "kind": "novel", + "stars_listed": 1, + "why": "Turn your existing SGLang / vLLM deployment into a Jev-compatible decision service. No training. No model changes. \u517c\u5bb9 Jev / System One \u7684\u8f7b\u91cf\u51b3\u7b56\u7f51\u5173\uff0c\u5c06\u6709\u9650\u5019\u9009\u4efb\u52a1\u8f6c\u6362\u4e3a next-token \u6253\u5206\uff0c\u76f4\u63a5\u590d\u7528 SGLang / vLLM\uff0c\u65e0\u9700\u8bad\u7ec3\u6216\u4fee\u6539\u6a21\u578b\u3002", + "html_url": "https://github.com/David-Lolly/Jev-Compatible", + "ok": true, + "github_id": 1379484833, + "description": "Turn your existing SGLang / vLLM deployment into a Jev-compatible decision service. No training. No model changes. \u628a\u4f60\u73b0\u6709\u7684 SGLang / vLLM \u90e8\u7f72\u53d8\u6210\u4e00\u4e2a\u517c\u5bb9 Jev \u7684\u51b3\u7b56\u670d\u52a1\u3002\u65e0\u9700\u4efb\u4f55\u4fee\u6539\u3002\u65e0\u9700\u8bad\u7ec3\u3002\u65e0\u9700\u66f4\u6539\u6a21\u578b\u3002", + "stars": 3, + "forks": 2, + "license": null, + "default_branch": "main", + "pushed_at": "2026-09-21T10:09:59Z", + "created_at": "2026-09-21T10:05:18Z", + "updated_at": "2026-09-21T10:55:56Z", + "size": 0, + "language": "Python", + "topics": [], + "archived": false, + "default_sha": "e52e963d85392fbf7a4490f2c58152eed4af4f45", + "empty": true, + "commit_message": "feat: add Jev-compatible vLLM backend\n\nAdd vLLM backend with logprob_token_ids capability probing, backend\nfactory, Docker deployment, benchmark data and docs (EN/zh-CN).\n\nCo-Authored-By: Claude Code ", + "readme_name": "README.md", + "readme_sha": "6e5ff22d40a69eb4e189af0d670e4c0c91edbeca", + "readme_size": 11573, + "readme_sha256": "f4f27fb24c93e5036a27dab21a6d8b0d1d177543c715d0ff30454c6e92ec2fbd", + "readme_preview": "# Jev Decision Service\n\n[English](README.md) | [\u4e2d\u6587](README.zh-CN.md)\n\nA lightweight Jev / System One compatible HTTP decision gateway.\n\nIt converts finite-candidate decision problems such as `noul`, `choice` and `score` into candidate token scoring tasks, directly reusing an already deployed [SGLang](https://github.com/sgl-project/sglang) or [vLLM](https://github.com/vllm-project/vllm) inference service.\n\nNo model training, no weight modifications, no inference engine modifications. The gateway itself loads no model, no PyTorch, no Transformers and no local tokenizer; it only communicates with the remote inference service over HTTP.\n\n\n## Why Use It\n\nFor tasks such as judgment, classification, routing and scoring, the answer itself comes from a finite set. There is no need to have the model generate a full JSON response, let alone produce a large amount of reasoning tokens for a simple decision.\n\nFor example:\n\n```text\nA \u2192 billing\nB \u2192 technical\nC \u2192 sales\n```\n\nJev Decision Service lets the model score the candidate answers at the next-token position:\n\n```text\nPrompt\n \u2193\nNext-token scoring\n \u2193\nA / B / C logprob\n \u2193\nSoftmax\n \u2193\nStructured decision result\n```\n\nIts main features:\n\n* Directly reuses existing SGLang / vLLM services; no changes to the model or inference engine.\n* No SFT, LoRA or extra classification head required \u2014 an ordinary generative LLM works directly as the decision model.\n* Each question requires only one next-token scoring call; no explanations, chain of thought, or full JSON generation.\n* The final JSON is constructed by the gateway from the probabilities, not by the model's output format.\n* Natively returns candidate probabilities, ready to use for threshold checks, human approval, agent routing, and more.\n* Compatible with Jev / System One `noul`, `choice` and `score` request and response structures.\n* Standard HTTP service, easily integrated into existing Python, Go, Java, Agent or workflow systems.\n* One-command Docker deployment; the gateway itself needs no GPU or model files.\n\n## Benchmark\n\nUsing exactly the same model and SGLang inference service, the same business decision task was called via ordinary Chat Completion and via Jev Decision Service.\n\nThe test task performs three judgments in one call:\n\n```text\nChoice: which team should handle it?\nNoul: does it need immediate action?\nScore: how severe is the issue?\n```\n\nTest environment:\n\n```text\nGPU: 4 \u00d7 NVIDIA L20\nModel: Qwen3.8-27B\nInference Engine: SGLang\nTensor Parallel: 4\n\nChat Completion:\nthinking = enabled\nstream = false\ntemperature = 0\n\nRuns: average of 3 runs\n```\n\nResults:\n\n| Metric | Chat Completion | Jev Decision Service |\n| ----------------- | --------------: | -------------------: |\n| Correct | 3/3 | 3/3 |\n| Valid JSON | 3/3 | 3/3 |\n| Valid Schema | 3/3 | 3/3 |\n| Avg Latency | 1282.3 ms | 232.1 ms |\n| Avg Input Tokens | 268 | 451 |\n| Avg Output Tokens | 212 | 3 |\n| Reasoning Tokens | 185 | 0 |\n\nOn this simple decision task, both approaches produced correct final answers.\n\nOrdinary Chat Completion generated on average 212 output tokens, of which 185 were reasoning tokens; Jev Decision Service generates no chain of thought and performs only one candidate scoring per question, producing on average only 3 scoring positions.\n\nFinal:\n\n```text\nChat Completion 1282.3 ms\nDecision Gateway 232.1 ms\n```\n\nIn this test environment, the end-to-end latency of the Decision Gateway is about `1/5.5` of Chat Completion.\n\nThe complete raw data of the benchmark is saved in:\n\n```text\nbenchmarks/results/20260921_174806/\n```\n\nThe directory contains, for every run:\n\n```text\nFull HTTP Request\nFull HTTP Response\nLatency\nToken Usage\nAnswer correctness\nJSON validity\nSchema validity\n```\n\nAll three runs are included in the averages; ", + "readme_len": 11492, + "description_hash": "347367770dbd" + }, + { + "id": "Dililianxice/jev-robotic-arm-benchmark", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "novel first-seen github peer", + "html_url": "https://github.com/Dililianxice/jev-robotic-arm-benchmark", + "ok": true, + "github_id": 1379533921, + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "default_branch": "main", + "pushed_at": "2026-09-21T10:43:27Z", + "created_at": "2026-09-21T10:43:26Z", + "updated_at": "2026-09-21T10:43:31Z", + "size": 0, + "language": null, + "topics": [], + "archived": false, + "default_sha": "2326e64ad5fb9b44d40b24dde2567479a2752aaa", + "empty": true, + "commit_message": "Initial commit", + "readme_name": "README.md", + "readme_sha": "6488066ce1214cac99ba37cfb1902a7b0186eba6", + "readme_size": 27, + "readme_sha256": "84d1449d9284d15ae2379a7cc8365446636bb1e6047271061b147a2a83ea1074", + "readme_preview": "# jev-robotic-arm-benchmark", + "readme_len": 27 + }, + { + "id": "JingHao-Leon/awesome-jev-apps", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "Jev \u4f18\u8d28\u5e94\u7528\u4e0e\u751f\u6001\u7cbe\u9009\uff5cSystem One \u51b3\u7b56\u6a21\u578b\uff1a\u5f00\u6e90\u5e94\u7528\u00b7SDK\u00b7\u5e73\u53f0\u96c6\u6210\u00b7\u5f00\u6e90\u590d\u523b\u00b7\u6559\u7a0b | curated apps & SDKs for TypeSafe AI's Jev model", + "html_url": "https://github.com/JingHao-Leon/awesome-jev-apps", + "ok": true, + "github_id": 1379494662, + "description": "Jev \u4f18\u8d28\u5e94\u7528\u4e0e\u751f\u6001\u7cbe\u9009\uff5cSystem One \u51b3\u7b56\u6a21\u578b\uff1a\u5f00\u6e90\u5e94\u7528\u00b7SDK\u00b7\u5e73\u53f0\u96c6\u6210\u00b7\u5f00\u6e90\u590d\u523b\u00b7\u6559\u7a0b | curated apps & SDKs for TypeSafe AI's Jev model", + "stars": 0, + "forks": 0, + "license": "MIT", + "default_branch": "main", + "pushed_at": "2026-09-21T10:13:26Z", + "created_at": "2026-09-21T10:13:05Z", + "updated_at": "2026-09-21T10:13:47Z", + "size": 0, + "language": null, + "topics": [ + "ai-agents", + "ai-routing", + "awesome-list", + "chinese", + "classification", + "jev", + "llm", + "structured-output", + "system-one", + "typesafe-ai" + ], + "archived": false, + "default_sha": "d3ef0254c4b23fb2b36748aef60d1c12043bc1a2", + "empty": true, + "commit_message": "feat: Jev \u4f18\u8d28\u5e94\u7528\u4e0e\u751f\u6001\u7cbe\u9009\u6e05\u5355\uff082026-09-21 \u9996\u4e2a\u5feb\u7167\uff09", + "readme_name": "README.md", + "readme_sha": "fd38a3c4ff2ecee3ede1aa5670177200b4d9fea5", + "readme_size": 16243, + "readme_sha256": "b9d3c9c0238cd94efa06dafb304acb0d778964f704a09673bbdece69d7d367b4", + "readme_preview": "
\n\n# Awesome Jev Apps\n\n**Jev\uff08TypeSafe AI\u300cSystem One\u300d\u51b3\u7b56\u6a21\u578b\uff09\u4f18\u8d28\u5e94\u7528\u4e0e\u751f\u6001\u7cbe\u9009 \u00b7 \u6301\u7eed\u66f4\u65b0**\n\n[![License: MIT](https://img.shields.io/badge/License-MIT-green)](LICENSE)\n[![Jev](https://img.shields.io/badge/Jev-jev--1.13.0-4F46E5)](https://docs.typesafe.ai)\n[![Hacker News](https://img.shields.io/badge/Hacker_News-1931_points-FF6600)](https://news.ycombinator.com/item?id=49717558)\n[![PRs Welcome](https://img.shields.io/badge/PRs-welcome-brightgreen)](#\u6536\u5f55\u6807\u51c6\u4e0e\u6295\u7a3f)\n[![\u751f\u6001\u5feb\u7167](https://img.shields.io/badge/\u751f\u6001\u5feb\u7167-2026.09.21-blue)](#\u5c40\u9650\u4e0e\u8bf4\u660e)\n\n
\n\n**Jev** \u662f TypeSafe AI \u4e8e 2026 \u5e74 9 \u6708 15 \u65e5\u53d1\u5e03\u7684\u9996\u4e2a **System One \u51b3\u7b56\u6a21\u578b**\uff08System One decision model\uff09\uff1a\u5b83\u4e0d\u50cf\u4f20\u7edf\u5927\u8bed\u8a00\u6a21\u578b\u90a3\u6837\u9010\u5b57\u751f\u6210\u6587\u672c\uff0c\u800c\u662f\u5bf9\u9884\u5148\u5b9a\u4e49\u597d\u7684\u95ee\u9898\u8fd4\u56de**\u5e26\u6821\u51c6\u6982\u7387\u7684\u7c7b\u578b\u5316\u7b54\u6848**\u2014\u2014\u4e00\u6b21\u8c03\u7528\u5e76\u884c\u5b8c\u6210\u5206\u7c7b\u3001\u8def\u7531\u3001\u6253\u5206\u4e0e\u662f\u975e\u5224\u65ad\u3002\u672c\u4ed3\u5e93\u6301\u7eed\u7cbe\u9009 Jev \u53d1\u5e03\u4ee5\u6765\u751f\u6001\u4e2d\u6d8c\u73b0\u7684**\u4f18\u8d28\u5f00\u6e90\u5e94\u7528\u3001SDK\u3001\u5e73\u53f0\u96c6\u6210\u3001\u5f00\u6e90\u590d\u523b\u4e0e\u6df1\u5ea6\u6559\u7a0b**\uff0c\u5e2e\u5de5\u7a0b\u5e08\u6700\u5feb\u627e\u5230\u300c\u628a Jev \u7528\u8d77\u6765\u300d\u7684\u53c2\u8003\u5b9e\u73b0\u3002\n\n> English: A curated, quality-first list of apps, SDKs, integrations, open replicas, and guides built with **Jev**, the first \"System One\" decision model by TypeSafe AI \u2014 typed, calibrated answers instead of generated text. 70\u2013500 ms latency, $0.042 per million input tokens, output free.\n\n## \u76ee\u5f55\n\n- [Jev \u662f\u4ec0\u4e48\uff0860 \u79d2\u7248\uff09](#jev-\u662f\u4ec0\u4e4860-\u79d2\u7248)\n- [30 \u79d2\u4e0a\u624b](#30-\u79d2\u4e0a\u624b)\n- [\u4f18\u8d28\u5f00\u6e90\u5e94\u7528](#\u4f18\u8d28\u5f00\u6e90\u5e94\u7528)\n- [SDK \u4e0e\u96c6\u6210](#sdk-\u4e0e\u96c6\u6210)\n- [\u5b98\u65b9\u8d44\u6e90](#\u5b98\u65b9\u8d44\u6e90)\n- [\u6df1\u5ea6\u6559\u7a0b\u4e0e\u8bc4\u6d4b](#\u6df1\u5ea6\u6559\u7a0b\u4e0e\u8bc4\u6d4b)\n- [\u793e\u533a\u8ba8\u8bba](#\u793e\u533a\u8ba8\u8bba)\n- [FAQ](#faq)\n- [\u6536\u5f55\u6807\u51c6\u4e0e\u6295\u7a3f](#\u6536\u5f55\u6807\u51c6\u4e0e\u6295\u7a3f)\n- [\u5c40\u9650\u4e0e\u8bf4\u660e](#\u5c40\u9650\u4e0e\u8bf4\u660e)\n\n## Jev \u662f\u4ec0\u4e48\uff0860 \u79d2\u7248\uff09\n\n| \u7ef4\u5ea6 | \u4f20\u7edf\u751f\u6210\u5f0f LLM | Jev\uff08System One\uff09 |\n| --- | --- | --- |\n| \u8f93\u51fa | \u81ea\u7531\u6587\u672c\uff08\u9700\u89e3\u6790\u6821\u9a8c\uff09 | \u9884\u5b9a\u4e49\u7c7b\u578b\u5316\u7b54\u6848 + \u6982\u7387 + \u7f6e\u4fe1\u5ea6 |\n| \u63d0\u95ee\u539f\u8bed | \u2014 | **choice**\uff08\u2264255 \u9009\u9879\uff09\u00b7 **score**\uff082\u201310 \u7ea7\u523b\u5ea6\uff09\u00b7 **noul**\uff080\u20131 \u662f\u5426\u6982\u7387\uff09 |\n| \u7aef\u5230\u7aef\u5ef6\u8fdf | 3\u2013329 \u79d2 | **70\u2013500 \u6beb\u79d2** |\n| \u5b9a\u4ef7 | \u8f93\u51fa\u4ef7\u7ea6\u4e3a\u8f93\u5165 5 \u500d | **\u8f93\u5165 $0.042 / \u767e\u4e07 token\uff0c\u8f93\u51fa\u514d\u8d39** |\n| \u64c5\u957f | \u5199\u4f5c\u3001\u751f\u6210\u3001\u590d\u6742\u63a8\u7406 | \u5206\u7c7b\u3001\u8def\u7531\u3001\u6253\u5206\u3001\u662f\u975e\u5224\u65ad\uff08\"\u4f1a\u601d\u8003\u7684\u667a\u80fd if \u8bed\u53e5\"\uff09 |\n\n- \u5b98\u65b9\u58f0\u79f0\u5728 System One \u7c7b\u4efb\u52a1\u4e0a\u6bd4\u53c2\u7167\u524d\u6cbf\u6a21\u578b**\u5feb 193.6 \u500d\u3001\u4fbf\u5b9c 444.6 \u500d**\uff08\u81ea\u5efa workflow evals\uff0c\u6279\u5224\u6027\u5206\u6790\u89c1 [TrueFoundry](https://www.truefoundry.com/blog/typesafe-ai-jev)\uff09\n- \u521b\u59cb\u4eba Diogo Almeida \u662f OpenAI \u524d\u7814\u7a76\u5458\u3001RLHF / InstructGPT \u8054\u5408\u53d1\u660e\u4eba\uff1b\u516c\u53f8\u79cd\u5b50\u8f6e $40M\uff08DCVC \u9886\u6295\uff09\n- \u5df2\u77e5\u5c40\u9650\uff1a\u4e0d\u4f1a\u7b97\u672f / \u8ba1\u6570 / \u65e5\u671f\u6bd4\u8f83\uff0c\u4e0d\u80fd\u751f\u6210\u6587\u672c\uff0c\u4e0d\u652f\u6301\u56fe\u7247\u4e0e\u97f3\u9891\uff1b\"\u96f6\u5e7b\u89c9\"\u6307**\u4e0d\u53ef\u80fd\u8fd4\u56de schema \u4e4b\u5916\u7684\u503c**\uff0c\u4e0d\u7b49\u4e8e\"\u6c38\u8fdc\u6b63\u786e\"\n\n## 30 \u79d2\u4e0a\u624b\n\n```bash\nnpm install @typesafe-ai/sdk\n```\n\n```js\nimport { choice, score, noul, TypeSafeClient } from \"@typesafe-ai/sdk\";\n\nconst client = new TypeSafeClient(); // \u9ed8\u8ba4\u6a21\u578b jev-latest\n\nconst r = await client.systemOne({\n state: { ticket: \"\u88ab\u91cd\u590d\u6263\u6b3e\uff0c\u8981\u6c42\u9000\u6b3e\", order: { id: \"A-104\", charges: [49, 49] } },\n questions: {\n department: choice(\"\u54ea\u4e2a\u56e2\u961f\u5904\u7406\", { billing: \"\u652f\u4ed8/\u8ba2\u9605\u95ee\u9898\", technical: \"\u6545\u969c/\u96c6\u6210\", other: \"\u5176\u4ed6\" }),\n urgency: score(\"\u7d27\u6025\u7a0b\u5ea6\", [\"\u4f4e\", \"\u4e2d\", \"\u9ad8\"]),\n refund: noul(\"\u662f\u5426\u5e94\u9000\u6b3e\"),\n },\n});\n// \u4e00\u6b21\u8bf7\u6c42\u5e76\u884c\u8fd4\u56de\u5168\u90e8\u7b54\u6848\uff0c\u6bcf\u4e2a\u90fd\u5e26\u6982\u7387\u4e0e\u7f6e\u4fe1\u5ea6\uff0c\u53ef\u6309\u7f6e\u4fe1\u5ea6\u5206\u6d41\u4eba\u5de5\n```\n\n## \u4f18\u8d28\u5f00\u6e90\u5e94\u7528\n\n> \u2605 \u6570\u4e3a 2026-09-21 \u5feb\u7167\uff0c\u968f\u65f6\u95f4\u53d8\u5316\u3002\n\n### Agent \u4e0e\u81ea\u52a8\u5316\n\n- [**browser-use/jev-ultrafast**](https://github.com/browser-use/jev-ultrafast) \u2014 \"i. am. speed.\"\uff1aJev \u8d1f\u8d23\u9009\u52a8\u4f5c\u3001\u53ea\u5728\u9700\u8981\u6253\u5b57\u65f6\u624d\u5524\u9192 LLM \u7684\u6d4f\u89c8\u5668 agent\uff0cGoogle Flights \u5b9e\u64cd\u6f14\u793a 7.1 \u79d2\uff08`Python`\uff0c13.8k\u2605\uff09\n- [**trycua/cua**](https://github.com/trycua/cua) \u2014 \u5f00\u6e90\u8ba1\u7b97\u673a\u64cd\u4f5c agent \u6846\u67b6\uff0c\u5185\u7f6e CUA-S1\uff1a\u8ba1\u7b97\u673a\u64cd\u4f5c\u51b3\u7b56\u4e13\u7528\u5c0f\u578b\u6a21\u578b\uff0825.4k\u2605\uff09\n- [**imanshu03/jev-browser-use**](https://github.com/imanshu03/jev-browser-use) \u2014 \u57fa\u4e8e Jev + CDP/Chromium \u7684\u8f7b\u91cf\u6d4f\u89c8\u5668\u81ea\u52a8\u5316\uff08`Python`\uff09\n\n### \u4ea4\u6613\u4e0e\u91d1\u878d\n\n- [**OpenByteInc/QuantDinger**](https://github.com/OpenByteInc/QuantDinger) \u2014 \u5f00\u6e90 AI \u4ea4\u6613\u64cd\u4f5c\u7cfb\u7edf\uff0c\u96c6\u6210 Jev System One\uff1a\u7b56\u7565\u7814\u7a76\u3001\u56de\u6d4b\u3001\u6a21\u62df/\u5b9e\u76d8\uff0c\u53ef\u642d\u591a\u79df\u6237\u4ea4\u6613 SaaS\uff08`Python`\uff0c11.9k\u2605\uff09\n- [**jarrodwatts/jev-trader**](https://github.com/jarrodwatts/jev-trader) \u2014 \u6bcf\u4e2a Monad \u533a\u5757\u505a\u4e00\u6b21 Jev \u4ea4\u6613\u51b3\u7b56\uff08Kuru MON-USDC\uff09\uff08`TypeScript`\uff0c1.7k\u2605\uff09\n- [**SaratAngajalaoffl/jeeva**](https://github.com/SaratAngajalaoffl/jeeva) \u2014 \u57fa\u4e8e Jev \u7684\u4e2d\u9891\u4ea4\u6613\u6846\u67b6\uff08`TypeScript`\uff09\n\n### \u5f00\u53d1\u8005\u5de5\u5177\n\n- [**tamaratran/fast-jev-compaction**](https://github.com/tamaratran/fast-jev-compaction) \u2014 Claude Code \u63d2\u4ef6\uff1a\u4e00\u6b21 Jev \u8bf7\u6c42\u7ed9\u5168\u90e8\u5de5\u5177\u8c03\u7528\u6253\u5206\u5e76\u538b\u7f29\u4e0a\u4e0b\u6587\uff0c\u4fdd\u7559\u5185\u5bb9\u9010\u5b57\u4e0d\u52a8\uff08`TypeScript`\uff0c5.7k\u2605\uff09\n- [**devagrawal09/jev-review**](https://github.com/devagrawal09/jev-review) \u2014 \u5206\u9636\u6bb5\u4ee3\u7801\u5ba1\u67e5\u5de5\u4f5c\u6d41 + \u672c\u5730\u4eea\u8868\u76d8\uff08`TypeScript`\uff0c438\u2605\uff09\n- [**dabit3/jev-experiments**](https://github.com/dabit3/jev-experiments) \u2014 \u5ef6\u8fdf\u654f\u611f\u573a\u666f\u7684 Jev demo \u5408\u96c6\uff08`TypeScript`\uff0c352\u2605\uff09\n- [**nozomi-koborinai/jev-spec**](https://github.com/nozomi-koborinai/jev-spec) \u2014 \u6bcf\u6b21 commit \u7528 Jev \u68c0\u67e5\u4ee3\u7801\u4e0e Markdown \u89c4\u683c\u6587\u6863\u7684\u6f02\u79fb\uff08`TypeScript`\uff09\n- [**maayanlevy/mysql-ailike**](https://github.com/maayanlevy/mysql-ailike) \u2014 MySQL \u63d2\u4ef6\uff1a\u6309\u81ea\u7136\u8bed\u8a00\u8bed\u4e49\u8fc7\u6ee4\u6570\u636e\u884c\uff08`C++`\uff09\n- [**lukstei/slop-grader", + "readme_len": 11777, + "description_hash": "bad6c97e201f" + }, + { + "id": "Keitark/jev-cats-and-dogs", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "novel first-seen github peer", + "html_url": "https://github.com/Keitark/jev-cats-and-dogs", + "ok": true, + "github_id": 1379475645, + "description": null, + "stars": 0, + "forks": 0, + "license": null, + "default_branch": "main", + "pushed_at": "2026-09-21T10:45:34Z", + "created_at": "2026-09-21T09:57:50Z", + "updated_at": "2026-09-21T10:45:38Z", + "size": 0, + "language": "Python", + "topics": [], + "archived": false, + "default_sha": "97e7db75dad07f4dc7d6505306e14bb06f9c6cd7", + "empty": true, + "commit_message": "Preview pencil sketch ASCII samples", + "readme_name": "README.md", + "readme_sha": "36fa0338d96cceca65ff7ce6058cd12e7129723d", + "readme_size": 5436, + "readme_sha256": "045ab99b66ea5e74c25c3ee32de70ad46c7f13830ba1c58773b5472b2c4b73f4", + "readme_preview": "# Jev Cats vs Dogs via 64x64 ASCII\n\nA small experiment for testing whether Jev can classify cat and dog images after the image is deliberately collapsed into plain text.\n\nThe core idea is intentionally simple:\n\n1. collect about 50 cat images and 50 dog images from Wikimedia Commons;\n2. center-crop each image to a square;\n3. convert it to grayscale;\n4. resize it to exactly 64 x 64 pixels;\n5. map brightness to ASCII characters;\n6. send only the resulting 64 x 64 character grid to Jev;\n7. force a binary choice: cat or dog;\n8. record accuracy, class-wise recall, probabilities, confidence, and latency.\n\nNo image bytes, filename, source URL, title, caption, or label are sent to Jev.\n\n## Why this is interesting\n\nThis is not a conventional vision benchmark. Jev receives a textual spatial representation generated from the image:\n\n~~~text\n ....::::---==++**##\n ...::::---===++***####\n ...:::----===+++***####%%\n...\n~~~\n\nSo the experiment probes whether spatial visual structure survives a very aggressive image-to-text bottleneck well enough for a general decision model to use it.\n\nThe 64 x 64 representation is 4096 visual characters plus line breaks. The character ramp is configurable so later runs can test whether performance depends on the amount of grayscale detail.\n\n## Setup\n\n~~~bash\ngit clone https://github.com/Keitark/jev-cats-and-dogs.git\ncd jev-cats-and-dogs\n\npython -m venv .venv\nsource .venv/bin/activate\n# Windows PowerShell: .venv\\Scripts\\Activate.ps1\n\npip install -r requirements.txt\ncp .env.example .env\n~~~\n\nSet your Jev key in .env:\n\n~~~dotenv\nJEV_API_KEY=your-key\nJEV_MODEL=jev-latest\nMODEL_TIMEOUT=30\n~~~\n\nTYPESAFE_API_KEY is also accepted as an alias for JEV_API_KEY.\n\n## 1. Collect about 100 images\n\n~~~bash\npython collect_images.py --per-class 50\n~~~\n\nThe collector uses Wikimedia Commons and stores:\n\n~~~text\ndata/\n raw/\n cat/\n dog/\n manifest.jsonl\n~~~\n\nThe manifest preserves the Commons source page, author/creator text, license metadata, and downloaded filename where available.\n\nThe raw dataset is intentionally gitignored. Re-run the collector to reproduce a fresh sample.\n\n## 2. Preview the exact ASCII input\n\n~~~bash\npython preview_ascii.py data/raw/cat/cat_0001.jpg\n~~~\n\nOr change the resolution / character ramp:\n\n~~~bash\npython preview_ascii.py data/raw/cat/cat_0001.jpg --width 32 --height 32\npython preview_ascii.py data/raw/cat/cat_0001.jpg --chars \"@%#*+=-:. \"\n~~~\n\nThe default ramp is:\n\n~~~text\n@%#*+=-:.\n~~~\n\nDark pixels map to dense characters and bright pixels map toward sparse visible characters. The default deliberately avoids spaces so trailing whitespace cannot be normalized away in transit.\n\n## 3. Run Jev classification\n\n~~~bash\npython benchmark.py --limit-per-class 50 --output results/jev_ascii_64x64.csv\n~~~\n\nThe default Jev request uses the same SystemOne style as the other Jev experiments:\n\n- endpoint: POST https://api.typesafe.ai/v1/systemone\n- state: instructions plus the 64 x 64 ASCII grid\n- question type: choice\n- choices: cat / dog\n\nThe benchmark prints a confusion matrix and:\n\n- overall accuracy\n- cat recall\n- dog recall\n- balanced accuracy\n- 95% Wilson interval for overall accuracy\n- mean latency\n- mean probability assigned to the correct class\n\nEach row of the CSV includes the true label, predicted label, p(cat), p(dog), confidence, latency, and local image path.\n\n## Useful experiments\n\nThe main run is 64 x 64, but the script keeps the representation configurable:\n\n~~~bash\npython benchmark.py --width 32 --height 32 --limit-per-class 50\npython benchmark.py --width 64 --height 64 --chars \"@#:. \"\npython benchmark.py --width 64 --height 64 --chars \"@ \"\npython benchmark.py --width 64 --height 64 --repeats 3\n~~~\n\nThat gives several useful ablations:\n\n- 64 x 64 vs 32 x 32: spatial resolution\n- 9-level vs reduced-level vs binary ASCII: grayscale information\n- repeated inference: decision stability\n- cat vs dog recall: class asymmetry\n\n## Input contract\n\nA Jev state looks ", + "readme_len": 5436 + }, + { + "id": "Manta-Boardgame/jev-chat", + "source": "github", + "kind": "revisit", + "stars_listed": 0, + "why": "description rewrite \u2014 Ask questions about your documents and get answers with probabilities. Windows & Android app for Jev (TypeSafe AI) via Vercel AI Gateway \u2014 reads whole books, PDFs, Word and images,", + "html_url": "https://github.com/Manta-Boardgame/jev-chat", + "ok": true, + "github_id": 1379433566, + "description": "Ask questions about your documents and get answers with probabilities. Windows & Android app for Jev (TypeSafe AI) via Vercel AI Gateway \u2014 reads whole books, PDFs, Word and images, with on-device OCR. Bring your own key.", + "stars": 0, + "forks": 0, + "license": null, + "default_branch": "main", + "pushed_at": "2026-09-21T10:53:16Z", + "created_at": "2026-09-21T09:25:25Z", + "updated_at": "2026-09-21T10:53:21Z", + "size": 100, + "language": "HTML", + "topics": [ + "ai-sdk", + "android", + "jev", + "ocr", + "pdf", + "vercel-ai-gateway", + "windows" + ], + "archived": false, + "default_sha": "44721bae8c2ee625bfc9d3876634899effaa928c", + "empty": false, + "commit_message": "Improve README and slim down the install\n\n- Rewrite README with highlights and per-platform sections\n- Document measured download and install sizes for Windows and Android\n- Add .npmrc (omit=optional)", + "readme_name": "README.md", + "readme_sha": "03272e4b9a9f24aa4f6acf4ef29ec5dbc701a137", + "readme_size": 6927, + "readme_sha256": "57dff06a65cac614b6644caa2ffd79a6aee2cf774ea49444e72ab278d03184b0", + "readme_preview": "# Jev Chat\n\n**Drop in a document, ask questions, get answers with probabilities \u2014 on Windows and Android.**\n\nJev Chat is a chat-style app for **Jev**, TypeSafe AI's typed decision model, running through **Vercel AI Gateway**.\nUnlike a chatbot, Jev doesn't write paragraphs. It gives you a decision \u2014 and tells you how sure it is.\n\n| Example question | Example answer |\n|---|---|\n| \"Does this contract allow early termination?\" | **Yes** \u2014 94% |\n| \"What is the tone of this review?\" | **Mixed** \u2014 positive 12%, mixed 84%, negative 4% |\n| \"How urgent is this support ticket?\" | **High** \u2014 low 2%, medium 10%, high 88% |\n\n> Unofficial, community-made app. Not affiliated with TypeSafe AI or Vercel.\n\n## Why Jev Chat\n\n**\ud83d\udcda Reads whole books, not just snippets.**\nJev accepts about 32,000 tokens per call. Jev Chat splits long documents automatically and merges the answers, so you can load a full novel or dissertation (up to 400,000 characters). In our test, it found a single sentence hidden near the end of a 130,000-character document with 98% confidence.\n\n**\ud83d\udcc4 Works with the files you actually have.**\nText, PDF, Word (`.docx`) and images. Scanned PDFs and photos are converted with on-device OCR. PDFs with Japanese/CJK fonts are handled correctly thanks to the bundled font maps.\n\n**\ud83c\udfaf Answers you can act on.**\nEvery answer comes with a probability for each option, so you can see when Jev is confident and when it is guessing. Ask many questions at once \u2014 one per line \u2014 and get them all back in one pass.\n\n**\ud83d\udd12 Private by design.**\nYour API key never leaves your device except to go to Vercel. OCR runs locally, so images are never uploaded. On Windows, the local server listens on `127.0.0.1` only.\n\n**\ud83d\udcb8 No surprise bills.**\nThe app shows the actual cost reported by Vercel and your remaining free credits, and it stops sending before you would be charged. Jev's list price at the time of writing is $0.042 per million input tokens, with free output.\n\n**\ud83d\udcf1 Same app, two platforms.**\nThe Windows and Android apps share the same interface. The Android app talks to Vercel directly \u2014 no PC needed.\n\n## Before you start: get your own API key\n\nThe app uses **your own** Vercel AI Gateway API key. Usage is billed to your own Vercel account.\n\n1. Create an account at [vercel.com](https://vercel.com/signup).\n2. Open **AI Gateway \u2192 API Keys \u2192 Create Key**. The key is shown only once, so save it.\n3. Free credits require adding a card. Adding a card alone does not charge you. Charges happen only if you buy credits or turn on auto top-up.\n4. Recommended: set a spending limit under **AI Gateway \u2192 Budgets** (for example $5).\n\nOn first launch the app asks for the key and stores it on your device only.\n\n## Platforms at a glance\n\n| | Windows | Android |\n|---|---|---|\n| Runs on | Windows 10 / 11 | Android 7.0+ |\n| Needs a PC to work | \u2014 | No |\n| OCR engine | Windows OCR (built in) | Google ML Kit (bundled, on-device) |\n| Scanned PDFs | Rendered by Windows | Rendered by Android |\n| Disk space | **About 60 MB** (+ Node.js, about 100 MB, if not installed yet) | **About 50 MB** |\n\n### App size in detail\n\nMeasured with version 1.0:\n\n| | Size | What it is |\n|---|---|---|\n| **Windows** \u2014 download (`git clone`) | 0.2 MB | Source code only |\n| **Windows** \u2014 after `npm install` | 59 MB | Vercel AI SDK and its dependencies (\u2248 18 MB) and pdf.js (\u2248 37 MB, of which 3.7 MB is actually used by the app) |\n| **Windows** \u2014 Node.js | \u2248 100 MB | Only if you don't have Node.js 22+ already |\n| **Android** \u2014 APK download | 50 MB | Most of it is the on-device text-recognition model |\n| **Android** \u2014 installed | 49 MB + under 1 MB of data | Grows only with your conversation history |\n\nMicrosoft Edge (used as the app window on Windows) and Windows OCR are part of Windows, so they add nothing.\n\n## Windows\n\n**Requirements:** Windows 10 or 11, [Node.js](https://nodejs.org/) 22 or later, Microsoft Edge (preinstalled).\n\n```bash\ngit clone https://github.com/Manta-Boardgame/jev-chat.git\ncd jev-chat\nnpm install\n``", + "readme_len": 6865, + "description_hash": "34abb6b7a9be" + }, + { + "id": "OriginalByteMe/system-one-chess-arena", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "A chess arena where Jev-powered agents compete, with live matches and decision-by-decision replays.", + "html_url": "https://github.com/OriginalByteMe/system-one-chess-arena", + "ok": true, + "github_id": 1379500465, + "description": "A chess arena where Jev-powered agents compete, with live matches and decision-by-decision replays.", + "stars": 0, + "forks": 0, + "license": null, + "default_branch": "main", + "pushed_at": "2026-09-21T10:17:39Z", + "created_at": "2026-09-21T10:17:32Z", + "updated_at": "2026-09-21T10:17:47Z", + "size": 0, + "language": "TypeScript", + "topics": [], + "archived": false, + "default_sha": "f8419ef995c1b030d93bd45741dd8d8a9cd76c9d", + "empty": true, + "commit_message": "avoid rebuilding boards for bracket reads", + "readme_name": "README.md", + "readme_sha": "9ef7fcae17574821945af5b20a15592fc0cd772b", + "readme_size": 1162, + "readme_sha256": "bb921571e369bd22e67d6d84a52a2f4a5aad49a61d50d197f41943d445ed479b", + "readme_preview": "# System One Chess Arena\n\nA chess league where AI decision systems compete as named players. Chess is the\nlaboratory: every turn has a closed set of legal choices, outcomes are\ndeterministic, and thousands of decisions run without human labelling. The\npoint is comparing decision architectures (typed decision models, LLMs,\nheuristics, engines) on strength, latency, cost and calibration, not beating\nStockfish.\n\nHosted entirely on Cloudflare: one Worker serves the UI and API, Durable\nObjects own season and game state, D1 holds the public read model.\n\nSee `PLAN.md` for the build order and `AGENTS.md`-adjacent vault note\n`Project Ideas/System One Chess Arena.md` for the full design rationale.\n\n## Development\n\n```sh\nbun install\nbun run typecheck\nbun run dev # wrangler dev on :8787\n```\n\n## Deploying\n\nThe site runs live on Cloudflare Workers, Durable Objects and D1. First-time\nsetup (creating the D1 database, applying migrations, secrets, Cloudflare\nAccess in front of `/admin`, Web Analytics) and every deploy after that is\ncovered start to finish in `DEPLOY.md`. Once set up:\n\n```sh\nbun run deploy # builds web/dist, then wrangler deploy\n```\n", + "readme_len": 1162, + "description_hash": "6a58e353fcea" + }, + { + "id": "Strernd/beer-jev", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "Should I have a(nother) beer? A tiny second opinion powered by Jev.", + "html_url": "https://github.com/Strernd/beer-jev", + "ok": true, + "github_id": 1379514461, + "description": "Should I have a(nother) beer? A tiny second opinion powered by Jev.", + "stars": 0, + "forks": 0, + "license": null, + "default_branch": "main", + "pushed_at": "2026-09-21T10:50:42Z", + "created_at": "2026-09-21T10:28:19Z", + "updated_at": "2026-09-21T10:50:46Z", + "size": 0, + "language": "TypeScript", + "topics": [], + "archived": false, + "default_sha": "cd83de1f0e53ced340233e4f9d4a3c31161e74de", + "empty": true, + "commit_message": "Add centered builder attribution", + "readme_name": "README.md", + "readme_sha": "e215bc4ccf138bbc38ad58ad57e92135484b3c0f", + "readme_size": 1450, + "readme_sha256": "60b55ff7df79af72590f9524208e46642bc32bdc175cdad41349681c0e2f958f", + "readme_preview": "This is a [Next.js](https://nextjs.org) project bootstrapped with [`create-next-app`](https://nextjs.org/docs/app/api-reference/cli/create-next-app).\n\n## Getting Started\n\nFirst, run the development server:\n\n```bash\nnpm run dev\n# or\nyarn dev\n# or\npnpm dev\n# or\nbun dev\n```\n\nOpen [http://localhost:3000](http://localhost:3000) with your browser to see the result.\n\nYou can start editing the page by modifying `app/page.tsx`. The page auto-updates as you edit the file.\n\nThis project uses [`next/font`](https://nextjs.org/docs/app/building-your-application/optimizing/fonts) to automatically optimize and load [Geist](https://vercel.com/font), a new font family for Vercel.\n\n## Learn More\n\nTo learn more about Next.js, take a look at the following resources:\n\n- [Next.js Documentation](https://nextjs.org/docs) - learn about Next.js features and API.\n- [Learn Next.js](https://nextjs.org/learn) - an interactive Next.js tutorial.\n\nYou can check out [the Next.js GitHub repository](https://github.com/vercel/next.js) - your feedback and contributions are welcome!\n\n## Deploy on Vercel\n\nThe easiest way to deploy your Next.js app is to use the [Vercel Platform](https://vercel.com/new?utm_medium=default-template&filter=next.js&utm_source=create-next-app&utm_campaign=create-next-app-readme) from the creators of Next.js.\n\nCheck out our [Next.js deployment documentation](https://nextjs.org/docs/app/building-your-application/deploying) for more details.\n", + "readme_len": 1450, + "description_hash": "9a765fc1ea7e" + }, + { + "id": "TianyuCodings/NanoJev", + "source": "github", + "kind": "revisit", + "stars_listed": 1680, + "why": "head SHA change 618cea6d\u219276fdfc9e \u2014 A nano replica of Jev: parallel decisions, dynamic candidates, and an end-to-end training pipeline.", + "html_url": "https://github.com/TianyuCodings/NanoJev", + "ok": true, + "github_id": 1374702542, + "description": "A nano replica of Jev: parallel decisions, dynamic candidates, and an end-to-end training pipeline.", + "stars": 1685, + "forks": 189, + "license": "MIT", + "default_branch": "main", + "pushed_at": "2026-09-21T09:58:28Z", + "created_at": "2026-09-17T16:08:30Z", + "updated_at": "2026-09-21T10:55:52Z", + "size": 64035, + "language": "Python", + "topics": [], + "archived": false, + "default_sha": "76fdfc9ecdca45a9bcef17991a07d3041a87685a", + "empty": false, + "commit_message": "Add JevHarness project link to READMEs", + "readme_name": "README.md", + "readme_sha": "a8f8afeb7e4462e45211aaeee391bfd4ee2e9cc2", + "readme_size": 9366, + "readme_sha256": "4536f6769de55391a83a2282632f5ad66f16966a7204cf3278242fe5d2892809", + "readme_preview": "# NanoJev \u2014 A nano replica of [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)\n\n**English** | [\u7b80\u4f53\u4e2d\u6587](README.zh-CN.md)\n\n**A 0.6B parallel decision model: states and questions in, complete probability distributions out. Zero output-token decoding.**\n\n> **New project: [JevHarness](https://github.com/TianyuCodings/JevHarness)** \u2014 Let an LLM build task-specific decision harnesses with [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev), with optional refinement using rewards and execution traces. Includes an interactive Pok\u00e9mon demo.\n\n[Play ViZDoom](https://nanojev-dev.tianyuchen99.chatgpt.site/?autoplay=1) \u00b7 [Maze & Snake](https://nanojev-dev.tianyuchen99.chatgpt.site/side-by-side?autoplay=1#maze) \u00b7 [Model](https://huggingface.co/C-Tianyu/NanoJev) \u00b7 [Dataset](https://huggingface.co/datasets/C-Tianyu/NanoJev-Data)\n\n**Now playing ViZDoom:** one shared checkpoint handles Basic aiming and Predict Position's moving-target rocket shots, alongside Maze and Snake.\n\n**4 tasks \u00b7 18,760 decision questions per data variant \u00b7 896 Predict Position expert episodes**\n\n## What's new\n\n**September 20, 2026 \u2014 One model, four games.**\n\n- **ViZDoom Basic:** **128/128** test successes, compared with **56/128** for [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev).\n- **ViZDoom Predict Position:** **27/128** test successes, up from **11/128** before this round; the matched [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) run also scores **11/128**. The policy learns when to turn, wait and fire at a moving target.\n- **16,333 ViZDoom questions** within an **18,760-question** mixed-task dataset per target variant, spanning train, dev, calibration, test and OOD.\n- **One model, four games:** the same step-400 checkpoint also completes the 50\u00d750 maze in **225 attempts** and collects **30 food items** during a full 256-step Snake run.\n\n## Three models, side by side\n\nReal browser replays of **[Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev), NanoJev and Untuned Qwen**. These animations loop automatically; click either one to open its interactive player. All four demos use the same current NanoJev checkpoint. The interactive development site currently requires access; all recordings can also be played locally using the commands below.\n\n### ViZDoom Basic \u00b7 Aim, then fire\n\n[![NanoJev eliminates the target with one shot while Jev and Untuned Qwen fail, shown side by side on the same game clock](assets/basic_unified_autoplay.gif)](https://nanojev-dev.tianyuchen99.chatgpt.site/?autoplay=1)\n\nMove into position, line up the target, fire. NanoJev eliminates the target with **one shot in 1.40 s**; [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) and Untuned Qwen each fire **19 shots** without an elimination before the deadline. The three panels share the same game clock and show original frames and action probabilities.\n\n### Find the exit \u00b7 50\u00d750 Maze\n\n[![Jev, current NanoJev and Untuned Qwen explore the same 50\u00d750 maze in the live three-panel viewer](assets/maze_unified_autoplay.gif)](https://nanojev-dev.tianyuchen99.chatgpt.site/side-by-side?autoplay=1#maze)\n\nNanoJev reaches the exit in **225 attempts**, versus **2,738** for [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) and **4,726** for Untuned Qwen. Each system combines local safety probabilities with the same exploration code and remembered open paths.\n\n[Play Snake](https://nanojev-dev.tianyuchen99.chatgpt.site/side-by-side?autoplay=1#snake) \u00b7 [Play Predict Position](https://nanojev-dev.tianyuchen99.chatgpt.site/predict-position?autoplay=1)\n\n## What NanoJev does\n\n- **Parallel decisions:** batch independent states, questions and candidate paths in one backbone forward.\n- **Dynamic candidates:** Choice returns a distribution over 2\u2013255 supplied candidates using a shared scoring head.\n- **Boolean and ordered scores:** predict a proposition's probability, or a distribution and e", + "readme_len": 9329, + "description_hash": "92ce9454eeeb" + }, + { + "id": "Tsagaanbayr1/jev-tetris", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "Real-time Tetris versus Jev, a TypeSafe decision model \u2014 spins, garbage, B2B chains, and decisions prefetched a piece ahead", + "html_url": "https://github.com/Tsagaanbayr1/jev-tetris", + "ok": true, + "github_id": 1376238451, + "description": "Real-time Tetris versus Jev, a TypeSafe decision model \u2014 spins, garbage, B2B chains, and decisions prefetched a piece ahead", + "stars": 0, + "forks": 0, + "license": null, + "default_branch": "main", + "pushed_at": "2026-09-21T10:46:48Z", + "created_at": "2026-09-18T17:52:25Z", + "updated_at": "2026-09-21T10:46:50Z", + "size": 51, + "language": "JavaScript", + "topics": [], + "archived": false, + "default_sha": "737cd8ae6f944db01845a90b563e743b0116b032", + "empty": false, + "commit_message": "Merge pull request #2 from Tsagaanbayr1/gaming-mode\n\nGaming mode: one-command launcher for Laya + game server", + "readme_name": "README.md", + "readme_sha": "ad28c5b592f10e9936aefd84dd6c601599d4a6f1", + "readme_size": 11089, + "readme_sha256": "97702b3a787d86a758bf7c141a2c28f42b822c6ad09854d0c2843fc37383b93c", + "readme_preview": "# Jev Tetris\n\nA real-time Tetris match in the browser against **Jev**, TypeSafe's decision\nmodel, or **Laya**, an open decision model running on your own machine \u2014 or\nwatch Jev and Laya play each other. Not a chat wrapper \u2014 a versus game with garbage, spins, combos\nand back-to-back chains, where every move the opponent makes is a decision the\nmodel makes, live, while the piece is falling.\n\nBoth games run in the browser in one animation loop, so your inputs never make a\nnetwork round trip. The server does one job: given a position, ask a model where\nthe piece goes.\n\n```\n\u250c\u2500 browser \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510 \u250c\u2500 server \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n\u2502 left game (you, or the Jev bot) \u2502 \u2502 reachability search \u2502\n\u2502 right game (Laya or Jev bot) \u2502\u25c0\u2500\u2500\u2500\u2500\u2500\u2500\u25b6\u2502 candidate shortlist \u2502\n\u2502 one animation frame drives both \u2502 /decide\u2502 one typed question \u2502\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518 ?model \u2514\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2518\n \u2502 \u2502\n /v1/systemone /v1/systemone\n \u2502 \u2502\n TypeSafe (Jev) ai/laya_server.py\n (Laya, local)\n```\n\n## Setup\n\n**Requirements:** Node 18+ (uses the built-in `fetch`, no npm dependencies). For\nLaya: Python 3.9+ and ~1.5 GB free disk; an Apple Silicon GPU (MPS) or CUDA is\nused automatically when present, otherwise CPU.\n\nYou can run either model alone or both. The server starts with whatever is\navailable, and the VS screen shows which models are online.\n\n### 1. Jev (TypeSafe API)\n\n```bash\ncp .env.example .env # then put your TypeSafe API key in .env\n```\n\n`.env` is git-ignored. The key stays on the server; the browser never sees it.\nWithout a key the server still runs, and Jev's moves fall back to the shortlist's\nown pick (labelled `FALLBACK` on screen).\n\n### 2. Laya (local, no key)\n\n[Laya Multilingual](https://huggingface.co/convaiinnovations/laya-multilingual)\n(Apache 2.0) runs in a small Python sidecar that serves the same `/v1/systemone`\nrequest shape as Jev:\n\n```bash\npython3 -m venv .venv-laya\n.venv-laya/bin/pip install laya\n.venv-laya/bin/python ai/laya_server.py # first run downloads the model (~640 MB)\n```\n\nIt listens on `127.0.0.1:8090` (`LAYA_PORT`, `LAYA_MODEL`, `LAYA_DEVICE` override\nit; the game server finds it via `LAYA_URL`). Leave it running.\n\n### 3. Start the game \u2014 gaming mode (one command)\n\n```bash\n./play.sh # or: npm run play\n```\n\nThis starts Laya and the game server in the background, waits until both are\nhealthy, and opens http://localhost:8081. It also sets up Laya's Python\nenvironment on first run if `.venv-laya` is missing. For speed (macOS, no admin\nrights needed):\n\n- both processes run at the highest throughput and latency QoS tiers\n (`taskpolicy -t 0 -l 0`), so they are never throttled as background work;\n- Laya uses every performance core for its CPU work and the GPU (MPS) for the\n model, with no cap on PyTorch's unified-memory use;\n- `caffeinate` keeps the Mac and display awake for as long as the server runs;\n- it warns if Low Power Mode is on or the Mac is on battery.\n\n```bash\n./play.sh --boost # also raise CPU priority to the maximum (sudo renice; asks for your password)\n./play.sh status # what is running, Laya's latency so far\n./play.sh stop # or: npm run stop\n```\n\nLogs go to `.run/laya.log` and `.run/server.log` (git-ignored). More RAM does\nnot make Laya faster: it needs ~1.3 GB and is limited by GPU compute.\n\nTo run just the game server in the foreground instead: `npm start`.\n\nAfter changing server-side code, restart `node server.js` \u2014 an old process on\n8081 keeps serving the old code (`pkill -f \"node server.js\"`).\n\n`npm test` runs the engine and AI test suites offline \u2014 no key, no model needed.\n\n## Playing\n\nC", + "readme_len": 10736, + "description_hash": "9833473756d4" + }, + { + "id": "Zafer-Liu/jev-xiangqi", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "Play Chinese Chess (Xiangqi) against Jev - TypeSafe System One decision model as the AI. Score fan-out over legal moves.", + "html_url": "https://github.com/Zafer-Liu/jev-xiangqi", + "ok": true, + "github_id": 1379514674, + "description": "Play Chinese Chess (Xiangqi) against Jev - TypeSafe System One decision model as the AI. Score fan-out over legal moves.", + "stars": 0, + "forks": 0, + "license": "BSD-2-Clause", + "default_branch": "main", + "pushed_at": "2026-09-21T10:36:39Z", + "created_at": "2026-09-21T10:28:29Z", + "updated_at": "2026-09-21T10:36:43Z", + "size": 0, + "language": "HTML", + "topics": [], + "archived": false, + "default_sha": "7181fab5bb27d0ad057d52e30b6f019496972b33", + "empty": true, + "commit_message": "chore: drop Railway deploy artifacts, keep as local demo\n\nRemove Procfile and revert server.py to the hardcoded local bind\n(127.0.0.1:8772), matching the sibling jev-demo-* projects. The app is\nnot de", + "readme_name": "README.md", + "readme_sha": "6bf367189d9ac63cb57109c0acf5fbc606a99c78", + "readme_size": 10400, + "readme_sha256": "1919ca5edd18acf2da616f1d8c59495a5729f4bf7c3314f1bdfac3a05f44c0ef", + "readme_preview": "# Jev Xiangqi\n\nPlay Chinese Chess against Jev. Every AI move is chosen by fanning out a Score 0-4 question per legal move in a single `system_one()` call \u2014 no minimax, no eval tables, just Jev's read of the position.\n\n*Built for [TypeSafe Jev](https://docs.typesafe.ai) \u2014 System One decision model.* | [\u4e2d\u6587](README.zh-CN.md)\n\n## How it plays\n\nOn the AI's turn, the frontend (xiangqi.js) enumerates every legal move and POSTs the position + move list to the backend. The backend builds one Jev call with:\n\n- **state**: ASCII board, FEN, side to move, check flag, last 12 moves, and a numbered list of legal moves (e.g. `[m7] h2e2 Red Cannon from h2 moves to e2`).\n- **questions**: one Score question per legal move (typically 30-45 in mid-game), each with the same 5 anchors \u2014 Blunder / Weak / Neutral / Good / Excellent.\n\nJev answers all N questions in parallel; the highest-scoring move is played. Ties go to the first move in Jev's natural order. The UI renders the top 10 with score bars so you can see the \"shape\" of Jev's evaluation, not just the pick.\n\n```\nYour move \u2500\u2500> xiangqi.js validates \u2500\u2500> POST /api/move\n \u00b7 state: FEN + ASCII + numbered legal moves\n \u00b7 questions: {m0..mN-1}: Score 0-4\n \u2193\n Jev (single call, parallel scoring)\n \u2193\n argmax score \u2500\u2500> AI plays \u2500\u2500> board updates\n```\n\n## Why Jev\n\nJev returns typed, calibrated answers with input priced at $0.042/MTok (output free). Measured on `jev-latest`, 2026-09-21, with the v2 rules-enhanced prompt:\n\n| Scenario | Legal moves | Input tokens | Cost | Wall latency |\n|---|---|---|---|---|\n| Opening (turn 1) | 44 | 7,676 | $0.000322 | 6.6 s |\n| Mid-game (turn 5) | 36 | 6,614 | $0.000278 | 6.6 s |\n\nMost of the latency is Jev scoring N questions in parallel; most of the tokens are the shared rules primer (~800) + per-question rubric. Cheap enough to iterate freely, slow enough that you'll want a coffee-friendly pace rather than blitz.\n\nThe interesting property isn't strength (Jev is not a chess engine \u2014 it will miss a 3-move tactic, and its top-2 margin is often under 0.1 points, making the pick near-random among the top few); it's that the same call gives you a *distribution*. You can see the top 10, gate on margin, or log the entropy without extra API calls. In our mid-game test it picked `e2e6` (\u70ae\u6253\u4e2d\u5352, cannon-jumps-pawn to take the central pawn) \u2014 a real opening line, not a random shuffle.\n\n## Prompt design (v2)\n\nThe naive \"score each move 0-4\" prompt plays weakly because Jev lacks domain grounding. v2 injects a shared primer into `state` (paid once per call, not per question):\n\n- **Rules**: piece movement, cannon jump-capture, palace/river constraints, flying-general rule, pawn river-crossing, win/draw conditions.\n- **Piece values**: Rook 9, Cannon 4.5, Knight 4, Adviser/Bishop 2, Pawn 1 (2 after crossing). Lets Jev evaluate trades numerically instead of vibes.\n- **Tactical motifs**: fork (\u6349\u53cc), discovered check (\u62bd\u5c06), cannon-behind-knight mate (\u9a6c\u540e\u70ae), smothered mate (\u95f7\u6740), iron-bolt (\u94c1\u95e8\u6813), hollow cannon (\u7a7a\u5934\u70ae), double check (\u53cc\u5c06).\n- **Positional principles**: center control, piece activation, king safety, crossed-pawn value, \"don't hang material\" with an explicit 3-step simulation instruction (your move \u2192 opponent's best reply \u2192 your answer).\n- **Local material count**: the backend parses the FEN and computes Red/Black material totals (with crossed-pawn bonus) so Jev doesn't have to count pieces from ASCII.\n\nPer-question instructions are shortened accordingly (rules live in state), so the net token delta vs v1 is roughly zero (+2% mid-game, \u22121% opening) while the decision quality improves.\n\n### v1 vs v2 on the same mid-game position (36 legal moves)\n\n| | v1 (naive) | v2 (rules-enhanced) |\n|---|---|---|\n| Top pick | `e2e6` (2.31) | `e2e6` (2.34) |\n| Margin over #2 | 0.46 | 0.5", + "readme_len": 10148, + "description_hash": "d248171f0d86" + }, + { + "id": "agrogov/jev-system-one-study", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "Jev System One Black-Box Study - Complete Reproducibility Bundle", + "html_url": "https://github.com/agrogov/jev-system-one-study", + "ok": true, + "github_id": 1379474272, + "description": "Jev System One Black-Box Study - Complete Reproducibility Bundle", + "stars": 0, + "forks": 0, + "license": null, + "default_branch": "main", + "pushed_at": "2026-09-21T10:03:15Z", + "created_at": "2026-09-21T09:56:43Z", + "updated_at": "2026-09-21T10:05:15Z", + "size": 0, + "language": "Python", + "topics": [], + "archived": false, + "default_sha": "2582528b6701f84500bfb300c58e94000fd6b7e8", + "empty": true, + "commit_message": "fix", + "readme_name": "README.md", + "readme_sha": "e1682d1d8de258f56178c95ef3eef553fa9d7160", + "readme_size": 1875, + "readme_sha256": "95e44ab7daa74b37ba82a9c2826db3e031edde6f11907343adf2225a70bce2d8", + "readme_preview": "# Jev System One Black-Box Study \u2014 Complete Reproducibility Bundle\n\nThis archive combines the complete research artifact in one self-contained tree:\n\n```text\njev-system-one-study/\n\u251c\u2500\u2500 README.md\n\u251c\u2500\u2500 benchmark/ # unified benchmark v1.0.0 (single codebase)\n\u251c\u2500\u2500 report/ # integrated research report and independent analysis\n\u2514\u2500\u2500 results/ # five self-contained recorded experiment folders used by the report\n```\n\n## Start here\n\n1. Read `report/JEV_SYSTEM_ONE_BLACKBOX_REPORT.md` for the scientific findings and reverse-engineered architecture.\n2. Read `benchmark/README.md` for Python setup, build, installation, suite documentation, exact reproduction commands, and output formats.\n3. Inspect `results/` for the complete raw API evidence (`raw.jsonl`, `cases.jsonl`, manifests, normalized data, reports, and figures).\n\n## Included study runs\n\n- `results/run_20260919T201050Z_core` \u2014 303-request concurrent architecture/robustness core run.\n- `results/run_20260919T201739Z_core` \u2014 26-request isolated core architecture scaling run.\n- `results/run_20260919T202139Z_full` \u2014 31-request isolated full boundary scaling run.\n- `results/run_20260919T203758Z_calibration_full` \u2014 5,810-request exact probabilistic calibration run.\n- `results/run_20260919T211437Z_semantic_full` \u2014 1,110-request semantic calibration run.\n\nThe `benchmark/` directory supersedes all earlier development benchmark versions and contains every experiment family used by these runs.\n\n## Recorded-run metadata\n\nEach directory under `results/` contains its own `manifest.json` with the exact benchmark configuration used for that run. The raw API evidence is preserved directly in each run's `raw.jsonl`; no duplicate legacy result archives are included.\n\nA prebuilt pure-Python wheel is also included under `benchmark/dist/`; source remains authoritative.\n", + "readme_len": 1839, + "description_hash": "020628cc44d0" + }, + { + "id": "baltzparra/jev-study", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "Estudo sobre Jev, o System One Model da TypeSafe AI: o que \u00e9, casos reais, onde \u00e9 relevante", + "html_url": "https://github.com/baltzparra/jev-study", + "ok": true, + "github_id": 1379475897, + "description": "Estudo sobre Jev, o System One Model da TypeSafe AI: o que \u00e9, casos reais, onde \u00e9 relevante", + "stars": 0, + "forks": 0, + "license": null, + "default_branch": "main", + "pushed_at": "2026-09-21T10:02:08Z", + "created_at": "2026-09-21T09:58:05Z", + "updated_at": "2026-09-21T10:02:12Z", + "size": 0, + "language": "HTML", + "topics": [], + "archived": false, + "default_sha": "17b6b4427fa6a72401f53b43f1ee74097fbbdc72", + "empty": true, + "commit_message": "Gr\u00e1ficos responsivos, separador de refer\u00eancias, override de tema\n\nCo-Authored-By: Claude Fable 5.1 ", + "readme_error": "gh: Not Found (HTTP 404)", + "empty_readme": true, + "description_hash": "889615b00d09" + }, + { + "id": "chrisns/homebrew-laya-mac-serve", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "Homebrew tap for Laya Serve", + "html_url": "https://github.com/chrisns/homebrew-laya-mac-serve", + "ok": true, + "github_id": 1379486865, + "description": "Homebrew tap for Laya Serve", + "stars": 0, + "forks": 0, + "license": "MIT", + "default_branch": "main", + "pushed_at": "2026-09-21T10:47:56Z", + "created_at": "2026-09-21T10:06:58Z", + "updated_at": "2026-09-21T10:48:00Z", + "size": 0, + "language": "Ruby", + "topics": [], + "archived": false, + "default_sha": "1b3c4c0bdb70a627e85fcc9d5be3fbd1e3449db2", + "empty": true, + "commit_message": "Drop the no-quarantine flag from the caveats\n\nHomebrew 7 removed that option.", + "readme_name": "README.md", + "readme_sha": "5e58082b7e9b8ff14a0f34de8929eedd0cf598ee", + "readme_size": 750, + "readme_sha256": "33472a323593320649755246e2eb5c435293e600974d7cffac3d478a5b7080a3", + "readme_preview": "# Homebrew tap for Laya Serve\n\n[Laya Serve](https://github.com/chrisns/laya-mac-serve) is a macOS menu bar\napplication. It gives n8n an OpenAI-compatible endpoint for the Laya classifier.\n\n## Install\n\n```bash\nbrew install --cask --no-quarantine chrisns/laya-mac-serve/laya-serve\n```\n\nLeave out `--no-quarantine` if you prefer to approve the application in System\nSettings. The application is ad-hoc signed, not notarised by Apple.\n\nThe download is about 1.1 GB. The application holds the model weights, so it needs no\nnetwork at run time.\n\n## Update\n\n```bash\nbrew update && brew upgrade --cask laya-serve\n```\n\nThe `Update Cask` workflow keeps this tap in step with each release of\n[chrisns/laya-mac-serve](https://github.com/chrisns/laya-mac-serve).\n", + "readme_len": 750, + "description_hash": "f05c26652dc6" + }, + { + "id": "devbackend/jevgo", + "source": "github", + "kind": "novel", + "stars_listed": 1, + "why": "Unofficial Go client for the TypeSafe AI System One API (Jev) \u2014 typed questions in, calibrated answers out.", + "html_url": "https://github.com/devbackend/jevgo", + "ok": true, + "github_id": 1379480941, + "description": "Unofficial Go client for the TypeSafe AI System One API (Jev) \u2014 typed questions in, calibrated answers out.", + "stars": 1, + "forks": 0, + "license": "MIT", + "default_branch": "main", + "pushed_at": "2026-09-21T10:05:45Z", + "created_at": "2026-09-21T10:02:15Z", + "updated_at": "2026-09-21T10:06:12Z", + "size": 0, + "language": "Go", + "topics": [ + "api-client", + "classification", + "go", + "golang", + "jev", + "llm", + "sdk", + "system-one", + "typesafe", + "typesafe-ai" + ], + "archived": false, + "default_sha": "849768d0830528e08fbd572c692a667f4d1854a9", + "empty": true, + "commit_message": "Initial release: Go SDK for the TypeSafe AI System One API\n\nUnofficial client for POST /v1/systemone and GET /v1/models with typed\nNoul/Choice/Score questions and answers, retries with backoff and\nRet", + "readme_name": "README.md", + "readme_sha": "40a6bec31e5ae6ff5a2e07392d0ff90a9e834cd6", + "readme_size": 5340, + "readme_sha256": "746232a1f92dec28e7dfc0479ad5f7e4e5de2f3d6d6157a0348009634f73b84b", + "readme_preview": "# jevgo\n\nGo client for the [TypeSafe AI](https://docs.typesafe.ai) System One API (`jev` models).\n\n> **Unofficial.** This is an independent community SDK. It is not affiliated with, endorsed by,\n> or supported by TypeSafe. \"TypeSafe\" and \"Jev\" are names of their respective owners.\n> For the official SDKs see [Python](https://github.com/typesafe-ai/typesafe-sdk-python)\n> and [JavaScript](https://github.com/typesafe-ai/typesafe-sdk-js).\n\nSend a `state` and a map of typed questions (Noul, Choice, Score), get typed answers back \u2014 in one call.\nThe library imports only the standard library (testify is used in tests only).\n\n```sh\ngo get github.com/devbackend/jevgo\n```\n\nRequires Go 1.24 or later.\n\n## Quick start\n\n```go\nclient, err := jevgo.New() // reads TYPESAFE_API_KEY\nif err != nil {\n\tlog.Fatal(err)\n}\n\nresp, err := client.SystemOne(ctx, jevgo.SystemOneRequest{\n\tState: \"Help! My payouts have been failing for 3 days.\",\n\tQuestions: jevgo.Questions{\n\t\t\"is_urgent\": jevgo.Noul(\"Does this convey urgency?\").\n\t\t\tWithCriteria(\"Explicitly time-sensitive\", \"No urgency expressed\"),\n\t\t\"department\": jevgo.Choice(\"Which team should handle this?\", map[string]any{\n\t\t\t\"billing\": \"Payments, invoicing, refunds\",\n\t\t\t\"technical\": \"Bugs, outages, integrations\",\n\t\t\t\"sales\": nil,\n\t\t}),\n\t\t\"frustration\": jevgo.ScoreOf(\"How frustrated is the customer?\", \"Calm\", \"Frustrated\", \"Very angry\"),\n\t},\n})\nif err != nil {\n\tlog.Fatal(err)\n}\n\nurgent, _ := resp.Answers.Noul(\"is_urgent\") // urgent.Noul == 0.95\ndept, _ := resp.Answers.Choice(\"department\") // dept.Choice, dept.Probabilities, dept.Confidence\nfrustration, _ := resp.Answers.Score(\"frustration\") // frustration.Score, .Level(), .Legend\n```\n\n`resp.Model` is the versioned model that answered (e.g. `jev-1.13.0`), `resp.Usage` holds token counts,\n`resp.RequestID` the `x-typesafe-request-id` header.\n\n## Questions\n\n| Constructor | Wire type | Answer |\n|---|---|---|\n| `Noul(instructions)` / `.WithCriteria(yes, no)` | `noul` | `NoulAnswer{Noul}` \u2014 probability of \"yes\" |\n| `Choice(instructions, map[string]any)` / `ChoiceOf(instructions, options...)` | `choice` | `ChoiceAnswer{Choice, Probabilities, Confidence}` |\n| `Score(instructions, levels...)` / `ScoreOf(instructions, levels...)` | `score` | `ScoreAnswer{Score, Legend, Probabilities, Confidence}` |\n\n`State`, instructions and every criteria description accept any JSON-encodable value \u2014 a string, map, slice\nor struct \u2014 so you can pass [structured](https://docs.typesafe.ai/primitives/advanced) data:\n\n```go\njevgo.Noul(map[string]any{\n\t\"potential_duplicate\": map[string]any{\"name\": \"John Smith\", \"location\": \"Oakland, California\"},\n\t\"question\": \"Is the resume for the same person as `potential_duplicate`?\",\n})\n```\n\nAnswers of a type this SDK version does not know decode into `UnknownAnswer` with the raw JSON,\nso a new server-side primitive never breaks existing code.\n\n## Configuration\n\n| Option | Env variable | Default |\n|---|---|---|\n| `WithAPIKey` | `TYPESAFE_API_KEY` | required |\n| `WithBaseURL` | `TYPESAFE_BASE_URL` | `https://api.typesafe.ai` |\n| `WithDefaultModel` | `TYPESAFE_DEFAULT_MODEL` | `jev-latest` |\n| `WithTimeout` | \u2014 | 10s per attempt |\n| `WithRetryPolicy` | \u2014 | `DefaultRetryPolicy()` |\n| `WithHTTPClient`, `WithHeaders`, `WithLogger` | \u2014 | \u2014 |\n\nExplicit options win over environment variables. Per-call overrides:\n`WithRequestTimeout`, `WithRequestRetryPolicy`, `WithRequestHeaders`.\n\n## Retries\n\n`DefaultRetryPolicy()` matches the official SDKs: 2 retries, exponential backoff 0.5s \u2192 5s with 25% jitter,\nretries on 408, 429 and 5xx (including `529 Overloaded`), connection errors and per-attempt timeouts,\nhonors `Retry-After` / `retry-after-ms` up to 60s, and stops once the total 30s budget would be exceeded.\n\n```go\npolicy := jevgo.DefaultRetryPolicy()\npolicy.MaxRetries = 5\nclient, _ := jevgo.New(jevgo.WithRetryPolicy(policy))\n\n// disable retries for one call\nclient.SystemOne(ctx, req, jevgo.WithRequestRetryPolicy(jevgo.RetryPolicy{}))\n`", + "readme_len": 5322, + "description_hash": "e928e8fefac4" + }, + { + "id": "emtay-com/fastlaya", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "Laya wrapped in fastapi", + "html_url": "https://github.com/emtay-com/fastlaya", + "ok": true, + "github_id": 1379481138, + "description": "Laya wrapped in fastapi", + "stars": 0, + "forks": 0, + "license": null, + "default_branch": "master", + "pushed_at": "2026-09-21T10:05:50Z", + "created_at": "2026-09-21T10:02:24Z", + "updated_at": "2026-09-21T10:02:29Z", + "size": 0, + "language": null, + "topics": [], + "archived": false, + "default_sha": "c314b99f4547a84c97e622f1f139e897511eecec", + "empty": true, + "commit_message": "Init commit", + "readme_name": "README.md", + "readme_sha": "5363adf64b536a7018059db12114100be42ee9d0", + "readme_size": 7030, + "readme_sha256": "66da948e0cdbca0982e3addc3d77265c6759ed2a8101dcff1da2672d32ba11a2", + "readme_preview": "# fastlaya\n\nServe [Laya](https://huggingface.co/convaiinnovations/laya) through FastAPI in one\ncontainer. Python 3.13, English and multilingual Laya checkpoints, and three identical\ndecision routes:\n\n- `POST /v1/systemone`\n- `POST /api/alpha/decisions`\n- `POST /ai/run`\n- `GET /health` \u2014 readiness and the actual inference device\n- `GET /docs` \u2014 interactive Swagger UI (`/openapi.json` is also available)\n\n## Run\n\n```sh\ndocker compose up --build -d --wait --wait-timeout 900\n# Or: make up\nmake test\n```\n\nOpen . The initial startup downloads approximately\n843 MB of checkpoint weights; allow several minutes and at least 8 GB available\nRAM. The CUDA-capable image itself is several GB. Weights persist in a Docker\nvolume. The service becomes ready only after loading succeeds. Compose creates\nthe `fastlaya` network, with service DNS name `fastlaya` and internal port `8000`.\n\n```sh\ncurl http://localhost:8000/v1/systemone \\\n -H 'Content-Type: application/json' \\\n --data-binary @examples/choice.json\n```\n\nAll routes accept the same `state` and `questions`. `state` can be text, an object,\nor a list. `model` accepts `laya` (English, the default when omitted) or\n`laya-multilingual` (including Dutch). Unknown model names return 422.\n`choice` takes at least two labeled\ncriteria (or a list of unique labels), `score` takes at least two ordered criteria,\nand `noul` optionally takes `true` and `false` criteria. Invalid requests return 422.\nThe three English examples and a Dutch example are in `examples/`.\nIn `/docs`, each decision endpoint has an **Examples** dropdown for the English\nchoice, noul, and score requests, plus the Dutch multilingual request. Select one\nand use **Try it out** to send it.\n\n`laya` loads at startup. `laya-multilingual` loads on its first request, then stays\nin memory for subsequent requests. Its first request may download about 678 MB\nof weights and tokenizer files if they are not cached. Both models share the\nconfigured cache and device selection. `/health` reports the default model's device.\n\n```sh\ncurl http://localhost:8000/v1/systemone \\\n -H 'Content-Type: application/json' \\\n --data-binary @examples/multilingual.json\n```\n\nResponses identify the selected checkpoint in `model` and preserve the SDK's\n`answers` and `usage` fields. Each named\nanswer contains `type`, `confidence`, and `action.act_probability`, along with:\n\n| Type | Result |\n| --- | --- |\n| `choice` | `choice` label and per-label `probabilities` |\n| `noul` | `noul`: probability of true, between 0 and 1 |\n| `score` | `score`: expected zero-based rubric index, `legend`, `probabilities` |\n\nRoute and request compatibility is provided; no separate Jev response conversion\nis applied. Predictions come from Laya, so smoke tests check valid answers and\nprobability ranges rather than requiring a particular classification.\n\n## CUDA and CPU\n\nThe default image contains CUDA 12.8 PyTorch wheels. `LAYA_DEVICE=auto` selects\nCUDA when `torch.cuda.is_available()` is true, otherwise CPU. Use `cpu`, `cuda`,\nor `cuda:N` to override. An explicitly requested unavailable device fails startup.\nLaya itself may fall back to CPU after an inference/device-memory error; `/health`\nreports its current device.\n\nOn NVIDIA hosts with the NVIDIA Container Toolkit, grant the container GPU access:\n\n```sh\ndocker compose -f compose.yaml -f compose.gpu.yaml up --build -d --wait --wait-timeout 900\n# Equivalent Make invocation:\nmake up COMPOSE='docker compose -f compose.yaml -f compose.gpu.yaml'\n```\n\nThis still runs one container. Docker cannot expose host GPUs from inside a\ncontainer; the override supplies `gpus: all`. Plain `docker run` uses `--gpus all`.\nFor a smaller CPU-only image: `TORCH_INDEX=cpu make up`. Rebuild with\n`TORCH_INDEX=cu128` before switching that image to GPU use.\n\n## Configuration and cache\n\nCopy `.env.example` to `.env` or export these variables:\n\n| Variable | Default | Meaning |\n| --- | --- | --- |\n| `HF_CACHE_DIR` | Persistent named volume in Compo", + "readme_len": 7026, + "description_hash": "eaed8add1d34" + }, + { + "id": "fajarhide/askgrep", + "source": "github", + "kind": "novel", + "stars_listed": 0, + "why": "grep for the questions you cannot write as a pattern. Reads every function instead of sampling a few. Powered by Jev, TypeSafe AI's System One model.", + "html_url": "https://github.com/fajarhide/askgrep", + "ok": true, + "github_id": 1379053838, + "description": "grep for the questions you cannot write as a pattern. Reads every function instead of sampling a few. Powered by Jev, TypeSafe AI's System One model.", + "stars": 0, + "forks": 0, + "license": "Apache-2.0", + "default_branch": "main", + "pushed_at": "2026-09-21T10:44:34Z", + "created_at": "2026-09-21T03:54:57Z", + "updated_at": "2026-09-21T10:44:37Z", + "size": 968, + "language": "Rust", + "topics": [ + "ai", + "classifier", + "cli", + "code-search", + "developer-tools", + "grep", + "jev", + "rust", + "security-scanning", + "static-analysis", + "system-one", + "typesafe" + ], + "archived": false, + "default_sha": "c89bb7553a9e8430a98da7f10bbe2bbff93d5c25", + "empty": false, + "commit_message": "Attribute Jev where a reader will actually see it\n\nThree places: under the tagline, in Backends, and a Built on section at the\nfoot. The page already leaned on Jev's price for its central claim, that\n", + "readme_name": "README.md", + "readme_sha": "900f5acfa98c1fe280a42248cb3ed47b1ecb709e", + "readme_size": 9838, + "readme_sha256": "5726a854d0d2d661aa46b45f67d5a24c424d7a12bea90b697b7ecbfc24feeb19", + "readme_preview": "# askgrep\n\ngrep for the questions you cannot write as a pattern.\n\n**Powered by [Jev](https://typesafe.ai), TypeSafe AI's System One model.** It\nreturns a calibrated probability instead of prose, which is what makes reading\nevery function affordable rather than sampling a few.\n\n![askgrep finding two SQL concatenations in a demo tree, then repeating the sweep for free from cache](media/demo.gif)\n\n```console\n$ askgrep \"builds an SQL query by concatenating a value that came from the request.\" demo/\norders.py:14 0.99 def search_orders(conn, request):\nreport.py:11 0.98 def top_products(conn, request):\n\n2 hits in 11 chunks | 0 cached | 3910 tokens | $0.0002\n```\n\nThat is `demo/` in this repo, so you can run the same line and get the same two\nhits. The other three queries there bind their parameters and are left alone.\n\nYou already know how to find `.unwrap()`. You do not know how to write the regex\nfor \"retries without backoff\", so today you guess a few patterns and hope.\n\n## What it gives an agent\n\nAsk a coding agent \"which functions build SQL from a request\" and it has two\noptions, both bad. Read the whole tree, or read some of it and guess.\n\nReading the whole tree is real money. The sweep below is 665,214 input tokens,\nand output tokens are free on a classifier and are not free on a chat model:\n\n| reading 1,866 functions with | input $/Mtok | this sweep |\n| --- | --- | --- |\n| askgrep | 0.042 | **$0.028** |\n| Claude Haiku 4.5 | 1.00 | $0.67 |\n| Claude Sonnet 5 | 2.00 | $1.33 |\n| Claude Opus 5 | 5.00 | $3.33 |\n\nSo the agent does the second thing. It greps a few patterns, opens a handful of\nfiles, and answers from those. You get an answer that sounds complete and has\nnever been checked against most of the codebase.\n\naskgrep turns that into a smaller job. It reads all 1,866 functions for under\nthree cents and hands back nine line numbers. The agent then opens nine\nfunctions instead of a thousand, with its context spent on the code that\nmatters rather than on everything that did not.\n\n```sh\naskgrep \"builds SQL by concatenating a value from a request\" --json src/ \\\n | jq -r 'select(.score > 0.9) | \"\\(.file):\\(.line)\"'\n```\n\nThe cheap model narrows, the expensive one acts. askgrep is only the first half,\nand it is the half nobody wants to pay frontier prices for.\n\n## It reads everything, not a sample\n\nThis is the part that matters. Tools built on an LLM have to sample, because\nreading every function costs real money, so they pick a few files by embedding\nsimilarity and answer from those. When the thing you were looking for sits in a\nfile that did not get picked, the answer is \"nothing found\" and you never learn\notherwise.\n\naskgrep asks about every chunk. A full sweep of a 1,866 function codebase costs\nunder three cents, so there is no reason to sample. Complete beats probably\ncomplete when the question is a security one.\n\n![every function in the demo tree scored in turn, the two that concatenate request values coming back at 0.99 and 0.98](media/sweep.gif)\n\nEvery row above was read and scored. The two that came back hot are the two that\nbuild SQL by concatenation. The [full 28 second\nversion](https://github.com/fajarhide/askgrep/releases/download/v0.1.0/askgrep-promo.mp4)\nis attached to the v0.1.0 release.\n\n## Install\n\n```sh\ncargo install --git https://github.com/fajarhide/askgrep\nexport TYPESAFE_API_KEY=... # https://typesafe.ai\n```\n\nOr take a binary from the [latest release](https://github.com/fajarhide/askgrep/releases):\nmacOS on Apple silicon or Intel, Linux on musl so it runs on any distro.\n\n`--dry-run` counts the chunks and prices the sweep without a key, so you can see\nwhat a run would cost before signing up for anything.\n\n## Backends\n\naskgrep needs one number per chunk, so anything that can produce one will do.\n\n```sh\naskgrep \"...\" # jev, if TYPESAFE_API_KEY is set\naskgrep \"...\" -b openai -m gpt-4o-mini # any OpenAI-compatible endpoint\naskgrep \"...\" -b openai --base-url", + "readme_len": 9838, + "description_hash": "c363afdbbdfb" + }, + { + "id": "hf:dataset:syvai/danish-dynaword-laya", + "source": "huggingface", + "kind": "revisit", + "hf_kind": "dataset", + "why": "HF card/tag rewrite \u2014 task_categories:text-classification, language:da, license:other, size_categories:100K [SEP]\n[MASK]