From 1f82e0b32de00933fb1ee930e93022bcce8c9be1 Mon Sep 17 00:00:00 2001 From: Cursor Agent Date: Sun, 20 Sep 2026 15:40:09 +0000 Subject: [PATCH] =?UTF-8?q?Post-#35=20hygiene:=20=C2=A7114=20citation,=20s?= =?UTF-8?q?cannable=20dumps,=20lock=20#34/#35,=20hop-ECE=20interleave?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Rebased onto latest main after merged #36/#38/#37/#40 (9824aa2). Do not reopen or amend #23–#40. Do not push onto open #41. Follow-up off main: ecosystem 0843 cited notes.md §113; README and CHANGELOG Unreleased still held dump walls; the 0843 lock ended at #33. Restore v0.4.0 scannability (dumps live in changelog-hourly). Extend the 0843 uniqueness lock through merged #34 and #35. Hop-ECE self-test covers reverse and even/odd interleave. uniqueness_gate checks Pages strings and refuses CHANGELOG/README dump walls. IDs unchanged: notes.md §114 / composition 289–302 / findings batch #97. Merged #31 still owns §113 / 273–288 / #96. Later folds keep §115–§118. Co-authored-by: Basit Mustafa <24601@users.noreply.github.com> --- .agents/skills/augustus/SKILL.md | 4 +- .../references/agent-self-assessment.md | 4 +- .../augustus/references/applied-mappings.md | 6 +- .../references/composition-algebra.md | 4 +- .agents/skills/augustus/references/faq.md | 6 +- .../augustus/references/formal-methods.md | 4 +- .../augustus/references/formal-semi-formal.md | 4 +- .../augustus/references/judgment-class.md | 4 +- .../skills/augustus/references/mappings.md | 4 +- .../augustus/references/mental-models.md | 8 +- .../augustus/references/methods-catalog.md | 4 +- .../augustus/references/mixed-architecture.md | 6 +- .../augustus/references/question-design.md | 4 +- .../augustus/references/toolbox-mapping.md | 4 +- .../skills/augustus/references/validation.md | 4 +- .../augustus/scripts/evaluate_decisions.py | 13 +- .../augustus/scripts/uniqueness_gate.py | 38 +++- CHANGELOG.md | 201 +++++------------- CONTRIBUTING.md | 9 +- README.md | 5 +- docs/ecosystem.md | 6 +- research/archive/findings.md | 6 +- research/changelog-hourly.md | 8 +- research/notes.md | 8 +- research/refresh-log.md | 12 +- 25 files changed, 160 insertions(+), 216 deletions(-) diff --git a/.agents/skills/augustus/SKILL.md b/.agents/skills/augustus/SKILL.md index 199ae9d..06aff66 100644 --- a/.agents/skills/augustus/SKILL.md +++ b/.agents/skills/augustus/SKILL.md @@ -243,7 +243,7 @@ requests skip the brainstorm and build. - **Empirical recipe**: worked on a stated dataset/model/version — retest. - **Hypothesis**: plausible, unestablished — label and test before relying. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 ## NanoJev unified-games-v1 densify (`notes.md` §115) @@ -265,4 +265,4 @@ User-provided 0915 uniqueness lock: TianyuCodings/NanoJev unified-games-v1 densi User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; 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 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/agent-self-assessment.md b/.agents/skills/augustus/references/agent-self-assessment.md index 1c4b64e..86bff78 100644 --- a/.agents/skills/augustus/references/agent-self-assessment.md +++ b/.agents/skills/augustus/references/agent-self-assessment.md @@ -929,7 +929,7 @@ Hourly 0843 agent self-assessment: omni-/ask-jev probabilities are advisory, not User-provided 0806 uniqueness lock: people who compare Jev against GPT-5.6 has never fine-tuned BERTForXYZ for living and it shows; zero shot classifiers; scale them as much as decoder only models; many problems solved with LLMs could have been solved with them, it was a skill issue; opt for DeBERTa and ModernBERT ones; BERTForXYZ → DeBERTa → ModernBERT; Jev vs GPT-5.6 bakeoffs are a category error; encoder / ZS classifiers; institutional HF voice; quote *theirs*; do not invent accuracy numbers; softmax/ZS scores still ≠ calibrated Noul; soft scores ≠ hard gates; @mervenoyann; likes 421 / 189; impressions 35498 / 9613; multimodal image<>text ZS as perception front-end; hf:MoritzLaurer/deberta-v3-large-zeroshot-v2.0 likes 139; hf:MoritzLaurer/ModernBERT-large-zeroshot-v2.0 likes 72; Bart, bert, deberta, modernbert, these are all LLMs; Maziyar quoted; Jev is exemplar not the mandate; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 User-provided 0915 uniqueness lock: TianyuCodings/NanoJev unified-games-v1 densify; A 0.6B parallel decision model: states and questions in, complete probability distributions out. Zero output-token decoding.; One model, four games; ViZDoom Basic 128/128 vs Jev 56/128; Predict Position 27/128 vs Jev 11/128; Maze 225 attempts vs Jev 2738; Snake 30 food / 256 steps; held-out Maze 4/10 Snake 8/8 Basic 128/128 Predict 27/128; Untuned Qwen3-0.6B baseline; 18,760 questions per variant; 16,333 ViZDoom; 896 Predict Position expert episodes; hard_lr1e5; mix weights 1/3, 1/3, 1/6, 1/6; Hub C-Tianyu/NanoJev revision unified-games-v1 likes 58; dataset C-Tianyu/NanoJev-Data likes 5; HEAD 618cea6d906d54e128360786d12f703fff2b1245; 1289★ / 158 forks / size 64035; README SHA 4190093c64ee75b26e9726daa3b00cbcf6d3157a; MIT; caijinchun/nanojev-arena ≠ liao96312/jev-arena-nanojev ≠ zwliJay/jev-forge ≠ NanoJev; not TypeSafe Jev; open replica / specialist gameplay S1; soft scores ≠ hard gates; Game success ≠ calibrated Noul; local type boolean ≠ TypeSafe noul; invented_signal false; do not reopen or amend PR #31/#32/#33/#35; notes.md §115 @@ -939,4 +939,4 @@ User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jc User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/applied-mappings.md b/.agents/skills/augustus/references/applied-mappings.md index 5aaa881..e4b1f77 100644 --- a/.agents/skills/augustus/references/applied-mappings.md +++ b/.agents/skills/augustus/references/applied-mappings.md @@ -2499,12 +2499,12 @@ vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases. xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; 不是 benchmark; 概率没做 calibration. Qwen2.5 ≠ Archer; Qwen 3.8 sparring ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; Archer still promised_not_landed. SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) lastModified UNCHANGED; Blackwood likes 2 gated manual; Hub archerhume/4rcherhume HTTP 401. -do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33. +do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35. Do not copy keys / `npm` / `pip` / `uv` / `powershell` / `.env` / `TYPESAFE_API_KEY`. Soft Noul ≠ hard safety. User-provided 0806 uniqueness lock: people who compare Jev against GPT-5.6 has never fine-tuned BERTForXYZ for living and it shows; zero shot classifiers; scale them as much as decoder only models; many problems solved with LLMs could have been solved with them, it was a skill issue; opt for DeBERTa and ModernBERT ones; BERTForXYZ → DeBERTa → ModernBERT; Jev vs GPT-5.6 bakeoffs are a category error; encoder / ZS classifiers; institutional HF voice; quote *theirs*; do not invent accuracy numbers; softmax/ZS scores still ≠ calibrated Noul; soft scores ≠ hard gates; @mervenoyann; likes 421 / 189; impressions 35498 / 9613; multimodal image<>text ZS as perception front-end; hf:MoritzLaurer/deberta-v3-large-zeroshot-v2.0 likes 139; hf:MoritzLaurer/ModernBERT-large-zeroshot-v2.0 likes 72; Bart, bert, deberta, modernbert, these are all LLMs; Maziyar quoted; Jev is exemplar not the mandate; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 **User-provided 0915 HIGH (`notes.md` §115).** NanoJev unified-games-v1 densify: A 0.6B parallel decision model: states and questions in, complete probability distributions out. Zero output-token decoding. One model, four games. ViZDoom Basic 128/128 vs Jev 56/128. held-out Maze 4/10 Snake 8/8 Basic 128/128 Predict 27/128. @@ -2518,4 +2518,4 @@ User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jc User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/composition-algebra.md b/.agents/skills/augustus/references/composition-algebra.md index a12619f..3950329 100644 --- a/.agents/skills/augustus/references/composition-algebra.md +++ b/.agents/skills/augustus/references/composition-algebra.md @@ -2273,8 +2273,8 @@ items 120–128 / 2041 items 111–119 / 1943 items 102–110 / mechanism / §60 six-gates / §78 v1.2 board. Soft Noul ≠ hard safety. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; 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 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/faq.md b/.agents/skills/augustus/references/faq.md index d970a99..6717be2 100644 --- a/.agents/skills/augustus/references/faq.md +++ b/.agents/skills/augustus/references/faq.md @@ -3625,7 +3625,7 @@ No. Everything runs locally in the browser. There is no framework, database, acc ## Did Archer land this hour? Reopen PR #23–#32? Push onto #31? -No, no, and no. Hub archerhume/4rcherhume HTTP 401. Archer still promised_not_landed. tracker likes 66 (+2 vs 64) lastModified UNCHANGED. Laya likes 861 (was 822). Blackwood likes 2 gated manual. SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506). do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33. Fresh PR off main. Open #31 is a separate 0743 fold. `invented_signal: false`. `notes.md` §113. +No, no, and no. Hub archerhume/4rcherhume HTTP 401. Archer still promised_not_landed. tracker likes 66 (+2 vs 64) lastModified UNCHANGED. Laya likes 861 (was 822). Blackwood likes 2 gated manual. SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506). do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35. Fresh PR off main. Open #31 is a separate 0743 fold. `invented_signal: false`. `notes.md` §113. User-provided 0806 uniqueness lock: people who compare Jev against GPT-5.6 has never fine-tuned BERTForXYZ for living and it shows; zero shot classifiers; scale them as much as decoder only models; many problems solved with LLMs could have been solved with them, it was a skill issue; opt for DeBERTa and ModernBERT ones; BERTForXYZ → DeBERTa → ModernBERT; Jev vs GPT-5.6 bakeoffs are a category error; encoder / ZS classifiers; institutional HF voice; quote *theirs*; do not invent accuracy numbers; softmax/ZS scores still ≠ calibrated Noul; soft scores ≠ hard gates; @mervenoyann; likes 421 / 189; impressions 35498 / 9613; multimodal image<>text ZS as perception front-end; hf:MoritzLaurer/deberta-v3-large-zeroshot-v2.0 likes 139; hf:MoritzLaurer/ModernBERT-large-zeroshot-v2.0 likes 72; Bart, bert, deberta, modernbert, these are all LLMs; Maziyar quoted; Jev is exemplar not the mandate; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29. @@ -3633,7 +3633,7 @@ User-provided 0806 uniqueness lock: people who compare Jev against GPT-5.6 has n Hourly 0646 uniqueness lock: 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. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 ## Collapse SemIf into a new OpenJev census? Is the rename a second species? @@ -3659,4 +3659,4 @@ No. SemIf ≠ kw2828/OpenJev playground ≠ zhihz/openjev ≠ apiplant/semif-rs User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; 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 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/formal-methods.md b/.agents/skills/augustus/references/formal-methods.md index f18ffa6..fcdcdae 100644 --- a/.agents/skills/augustus/references/formal-methods.md +++ b/.agents/skills/augustus/references/formal-methods.md @@ -1362,10 +1362,10 @@ Related: `formal-semi-formal.md` (one screen); `mental-models.md`; `composition-algebra.md` positions 3 and 9; `mixed-architecture.md` preference lint; `faq.md`; `boundary-audit.md`. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; 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 User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/formal-semi-formal.md b/.agents/skills/augustus/references/formal-semi-formal.md index e730be9..6709a10 100644 --- a/.agents/skills/augustus/references/formal-semi-formal.md +++ b/.agents/skills/augustus/references/formal-semi-formal.md @@ -64,7 +64,7 @@ leftovers (`mappings.md` §18). Putting the Noul first is the rejected design. Hourly 0843: calibration does not compose. hop-ECE is a SENSOR with zero power on clustered regime errors. Deferred Crispification. Do not launder hop-ECE as a trajectory proof. Qwen/Qwen3.8-27B ≠ Archer. Soft Noul ≠ hard safety. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 **NanoJev unified-games-v1 (`notes.md` §115).** Gameplay success is a SENSOR, not a proof. local type boolean ≠ @@ -81,4 +81,4 @@ User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jc User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/judgment-class.md b/.agents/skills/augustus/references/judgment-class.md index 0c7fa2b..67b02fb 100644 --- a/.agents/skills/augustus/references/judgment-class.md +++ b/.agents/skills/augustus/references/judgment-class.md @@ -1456,7 +1456,7 @@ Hourly 0743 uniqueness lock: Heman10x-NGU/Verdict-open-jev ≠ Heman10x-NGU/open User-provided 0806 uniqueness lock: people who compare Jev against GPT-5.6 has never fine-tuned BERTForXYZ for living and it shows; zero shot classifiers; scale them as much as decoder only models; many problems solved with LLMs could have been solved with them, it was a skill issue; opt for DeBERTa and ModernBERT ones; BERTForXYZ → DeBERTa → ModernBERT; Jev vs GPT-5.6 bakeoffs are a category error; encoder / ZS classifiers; institutional HF voice; quote *theirs*; do not invent accuracy numbers; softmax/ZS scores still ≠ calibrated Noul; soft scores ≠ hard gates; @mervenoyann; likes 421 / 189; impressions 35498 / 9613; multimodal image<>text ZS as perception front-end; hf:MoritzLaurer/deberta-v3-large-zeroshot-v2.0 likes 139; hf:MoritzLaurer/ModernBERT-large-zeroshot-v2.0 likes 72; Bart, bert, deberta, modernbert, these are all LLMs; Maziyar quoted; Jev is exemplar not the mandate; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 User-provided 0915 uniqueness lock: TianyuCodings/NanoJev unified-games-v1 densify; A 0.6B parallel decision model: states and questions in, complete probability distributions out. Zero output-token decoding.; One model, four games; ViZDoom Basic 128/128 vs Jev 56/128; Predict Position 27/128 vs Jev 11/128; Maze 225 attempts vs Jev 2738; Snake 30 food / 256 steps; held-out Maze 4/10 Snake 8/8 Basic 128/128 Predict 27/128; Untuned Qwen3-0.6B baseline; 18,760 questions per variant; 16,333 ViZDoom; 896 Predict Position expert episodes; hard_lr1e5; mix weights 1/3, 1/3, 1/6, 1/6; Hub C-Tianyu/NanoJev revision unified-games-v1 likes 58; dataset C-Tianyu/NanoJev-Data likes 5; HEAD 618cea6d906d54e128360786d12f703fff2b1245; 1289★ / 158 forks / size 64035; README SHA 4190093c64ee75b26e9726daa3b00cbcf6d3157a; MIT; caijinchun/nanojev-arena ≠ liao96312/jev-arena-nanojev ≠ zwliJay/jev-forge ≠ NanoJev; not TypeSafe Jev; open replica / specialist gameplay S1; soft scores ≠ hard gates; Game success ≠ calibrated Noul; local type boolean ≠ TypeSafe noul; invented_signal false; do not reopen or amend PR #31/#32/#33/#35; notes.md §115 @@ -1464,4 +1464,4 @@ User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jc User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/mappings.md b/.agents/skills/augustus/references/mappings.md index 0044046..06b3119 100644 --- a/.agents/skills/augustus/references/mappings.md +++ b/.agents/skills/augustus/references/mappings.md @@ -2464,7 +2464,7 @@ Hourly 0843 mappings (`notes.md` §114): hysteresis policy latch; instruct-tunin User-provided 0806 uniqueness lock: people who compare Jev against GPT-5.6 has never fine-tuned BERTForXYZ for living and it shows; zero shot classifiers; scale them as much as decoder only models; many problems solved with LLMs could have been solved with them, it was a skill issue; opt for DeBERTa and ModernBERT ones; BERTForXYZ → DeBERTa → ModernBERT; Jev vs GPT-5.6 bakeoffs are a category error; encoder / ZS classifiers; institutional HF voice; quote *theirs*; do not invent accuracy numbers; softmax/ZS scores still ≠ calibrated Noul; soft scores ≠ hard gates; @mervenoyann; likes 421 / 189; impressions 35498 / 9613; multimodal image<>text ZS as perception front-end; hf:MoritzLaurer/deberta-v3-large-zeroshot-v2.0 likes 139; hf:MoritzLaurer/ModernBERT-large-zeroshot-v2.0 likes 72; Bart, bert, deberta, modernbert, these are all LLMs; Maziyar quoted; Jev is exemplar not the mandate; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 **0915 HIGH (`notes.md` §115):** TianyuCodings/NanoJev unified-games-v1 densify; A 0.6B parallel decision model: states and questions in, complete probability distributions out. Zero output-token decoding.; One model, four games; ViZDoom Basic 128/128 vs Jev 56/128; Predict Position 27/128 vs Jev 11/128; Maze 225 attempts vs Jev 2738; Snake 30 food / 256 steps; held-out Maze 4/10 Snake 8/8 Basic 128/128 Predict 27/128; Untuned Qwen3-0.6B baseline; not TypeSafe Jev; open replica / specialist gameplay S1; Game success ≠ calibrated Noul; local type boolean ≠ TypeSafe noul; caijinchun/nanojev-arena ≠ liao96312/jev-arena-nanojev ≠ zwliJay/jev-forge ≠ NanoJev; do not reopen or amend PR #31/#32/#33/#35; notes.md §115 User-provided 0915 uniqueness lock: TianyuCodings/NanoJev unified-games-v1 densify; A 0.6B parallel decision model: states and questions in, complete probability distributions out. Zero output-token decoding.; One model, four games; ViZDoom Basic 128/128 vs Jev 56/128; Predict Position 27/128 vs Jev 11/128; Maze 225 attempts vs Jev 2738; Snake 30 food / 256 steps; held-out Maze 4/10 Snake 8/8 Basic 128/128 Predict 27/128; Untuned Qwen3-0.6B baseline; 18,760 questions per variant; 16,333 ViZDoom; 896 Predict Position expert episodes; hard_lr1e5; mix weights 1/3, 1/3, 1/6, 1/6; Hub C-Tianyu/NanoJev revision unified-games-v1 likes 58; dataset C-Tianyu/NanoJev-Data likes 5; HEAD 618cea6d906d54e128360786d12f703fff2b1245; 1289★ / 158 forks / size 64035; README SHA 4190093c64ee75b26e9726daa3b00cbcf6d3157a; MIT; caijinchun/nanojev-arena ≠ liao96312/jev-arena-nanojev ≠ zwliJay/jev-forge ≠ NanoJev; not TypeSafe Jev; open replica / specialist gameplay S1; soft scores ≠ hard gates; Game success ≠ calibrated Noul; local type boolean ≠ TypeSafe noul; invented_signal false; do not reopen or amend PR #31/#32/#33/#35; notes.md §115 @@ -2473,4 +2473,4 @@ User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jc User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/mental-models.md b/.agents/skills/augustus/references/mental-models.md index f9e8885..dc05b04 100644 --- a/.agents/skills/augustus/references/mental-models.md +++ b/.agents/skills/augustus/references/mental-models.md @@ -2820,8 +2820,8 @@ Mathematical / logical / algorithmic mental models for Jev-class categorization/scoring across AI / SWE / business / knowledge work / life — not SWE-only. Do **not** re-fold §112 / §111 / §109 tunahan census as a -sibling. Fresh PR off main; never reopen merged #7–**#33**. -do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33. +sibling. Fresh PR off main; never reopen merged #7–**#35**. +do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35. Do **not** reopen or amend merged #31. Skip Archer rewrite (still **promised_not_landed**). Quote *theirs*. Soft Noul ≠ hard safety. Formal methods **compose** with scoring; a @@ -2968,8 +2968,8 @@ Related: `mappings.md` §1–§18, `methods-catalog.md`, `toolbox-mapping.md`, `composition-algebra.md`, `formal-methods.md`, `formal-semi-formal.md`, `faq.md`. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; 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 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/methods-catalog.md b/.agents/skills/augustus/references/methods-catalog.md index d9d999e..942f5e3 100644 --- a/.agents/skills/augustus/references/methods-catalog.md +++ b/.agents/skills/augustus/references/methods-catalog.md @@ -267,7 +267,7 @@ calibration), you don't yet have a mapping — you have a metaphor. - Program synthesis candidate ranking — fan-out nouls as the ranker, kept to shortlists (heuristics→rerank shape). **Hypothesis.** -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 **0920 jcr:** geometric-mean path score is a control statistic (cookbook beam cousin), not a calibrated Noul and not a permission bit. `notes.md` §116. @@ -275,4 +275,4 @@ User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jc User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/mixed-architecture.md b/.agents/skills/augustus/references/mixed-architecture.md index 0ce9178..42d1fdb 100644 --- a/.agents/skills/augustus/references/mixed-architecture.md +++ b/.agents/skills/augustus/references/mixed-architecture.md @@ -485,7 +485,7 @@ Hourly 0743 uniqueness lock: Heman10x-NGU/Verdict-open-jev ≠ Heman10x-NGU/open | Treat equal-width ECE as the ECE / quote jeval drift as a captured run / collapse into jevals | **Fail closed** (binning is a decision; M2 specified) | pd.cut bins by equal width while jeval bins by quantile. ECE 0.113 and ECE 0.076. jeval drift is not implemented yet. rlaope/jeval ≠ dayhaysoos/jevals. `notes.md` §114 | | Treat hop-ECE as a trajectory audit / quote 25–60× / treat Qwen 3.8 as Archer | **Fail closed** (calibration does not compose; headline withdrawn) | ECE has exactly zero statistical power to detect the failure mode that kills trajectories. 25–60× headline withdrawn. P(all-correct): 0.0071 vs 0.0001. TCE / AMS. Deferred Crispification. Qwen 3.8 sparring ≠ Archer. `notes.md` §114 | | Mint g0runmezadam as a second 947-repo census / treat A/B/C as Harbor | **Fail closed** (rename densify) | g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307). 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56. catalog ≠ endorsement. `notes.md` §114 | -| Hard-gate ask-jev p / treat 0.5 screening as 100% sensitivity / treat BERT-Base 93.02 as zero-shot / treat circuit-vl-4b as Archer / treat xiaohuaxi as a benchmark / reopen #23–#32 / push onto #31 | **Fail closed** (advisory / theater / category error / prior PRs) | 13 focused checks and one mutually exclusive outcome. Probabilities are advisory, not calibrated guarantees. omni-/ask-jev ≠ pedroknigge/mcp_jev. light_cutoff_applied_to_combination 0. BANKING77 Accuracy BERT-Base 93.02 Jev 79.90. BERT figures are published supervised references, not zero-shot. Analyse jev calibration (NLL, ECE) backlog. ranking ≠ calibration. AND: product (independence assumed and recorded in the trace). chat model's stated confidence is not calibrated. circuit-vl-4b ≠ Archer. 不是 benchmark; 概率没做 calibration; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study. Bring your own API key. recorded run, kinematic animation. 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません. vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases. do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33. `notes.md` §114 | +| Hard-gate ask-jev p / treat 0.5 screening as 100% sensitivity / treat BERT-Base 93.02 as zero-shot / treat circuit-vl-4b as Archer / treat xiaohuaxi as a benchmark / reopen #23–#32 / push onto #31 | **Fail closed** (advisory / theater / category error / prior PRs) | 13 focused checks and one mutually exclusive outcome. Probabilities are advisory, not calibrated guarantees. omni-/ask-jev ≠ pedroknigge/mcp_jev. light_cutoff_applied_to_combination 0. BANKING77 Accuracy BERT-Base 93.02 Jev 79.90. BERT figures are published supervised references, not zero-shot. Analyse jev calibration (NLL, ECE) backlog. ranking ≠ calibration. AND: product (independence assumed and recorded in the trace). chat model's stated confidence is not calibrated. circuit-vl-4b ≠ Archer. 不是 benchmark; 概率没做 calibration; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study. Bring your own API key. recorded run, kinematic animation. 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません. vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases. do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35. `notes.md` §114 | | Treat tracker likes 66 as Archer landing | **Fail closed** (likes ephemeral; lastModified UNCHANGED) | tracker likes 66 (+2 vs 64) lastModified UNCHANGED. Laya likes 861 (was 822). Blackwood likes 2 gated manual. Archer still promised_not_landed. Hub archerhume/4rcherhume HTTP 401. SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506). `notes.md` §114 | | Treat llm-to-jev as Jev / treat the SDK export as equivalent behavior / skip human review | **Fail closed** (conversion assistant ≠ gate) | Turn decision-shaped LLM prompts into proposed Jev primitives. This is a conversion assistant, not an automatic guarantee of equivalent behavior. Generated instructions and criteria must be reviewed before production use. `notes.md` §118 | @@ -1436,10 +1436,10 @@ Live docs + typesafe-ai skill revision actually used (when the family is Jev): Propose three *placements* (prefilter vs post-verify vs replace-this-one- classifier-step), recommend one, and name the experiment that kills it. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; 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 User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/question-design.md b/.agents/skills/augustus/references/question-design.md index 00f2c4b..3b5dc7e 100644 --- a/.agents/skills/augustus/references/question-design.md +++ b/.agents/skills/augustus/references/question-design.md @@ -416,10 +416,10 @@ Hourly 0541 uniqueness lock: calibration beyond ~500 tokens unmeasured; 11.57s v Confidence-gated routing (doc defaults): act / confirm-or-flag / hand off, floors 0.5–0.6, 0.85–0.9 for high-stakes — starting points, not constants (calibration must be re-measured per dataset; see validation.md). -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; 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 User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/toolbox-mapping.md b/.agents/skills/augustus/references/toolbox-mapping.md index 238f469..8d8837f 100644 --- a/.agents/skills/augustus/references/toolbox-mapping.md +++ b/.agents/skills/augustus/references/toolbox-mapping.md @@ -341,7 +341,7 @@ importance sampling) and redo step 3 of the sweep on it. The library of mappings grows one falsified card at a time; a card enters mappings.md only with an acceptance test that ran. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 User-provided 0915 uniqueness lock: TianyuCodings/NanoJev unified-games-v1 densify; A 0.6B parallel decision model: states and questions in, complete probability distributions out. Zero output-token decoding.; One model, four games; ViZDoom Basic 128/128 vs Jev 56/128; Predict Position 27/128 vs Jev 11/128; Maze 225 attempts vs Jev 2738; Snake 30 food / 256 steps; held-out Maze 4/10 Snake 8/8 Basic 128/128 Predict 27/128; Untuned Qwen3-0.6B baseline; 18,760 questions per variant; 16,333 ViZDoom; 896 Predict Position expert episodes; hard_lr1e5; mix weights 1/3, 1/3, 1/6, 1/6; Hub C-Tianyu/NanoJev revision unified-games-v1 likes 58; dataset C-Tianyu/NanoJev-Data likes 5; HEAD 618cea6d906d54e128360786d12f703fff2b1245; 1289★ / 158 forks / size 64035; README SHA 4190093c64ee75b26e9726daa3b00cbcf6d3157a; MIT; caijinchun/nanojev-arena ≠ liao96312/jev-arena-nanojev ≠ zwliJay/jev-forge ≠ NanoJev; not TypeSafe Jev; open replica / specialist gameplay S1; soft scores ≠ hard gates; Game success ≠ calibrated Noul; local type boolean ≠ TypeSafe noul; invented_signal false; do not reopen or amend PR #31/#32/#33/#35; notes.md §115 @@ -351,4 +351,4 @@ User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jc User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/references/validation.md b/.agents/skills/augustus/references/validation.md index 18e2e70..be81c38 100644 --- a/.agents/skills/augustus/references/validation.md +++ b/.agents/skills/augustus/references/validation.md @@ -1149,10 +1149,10 @@ User-provided 0806 uniqueness lock: people who compare Jev against GPT-5.6 has n **0940 HIGH (`notes.md` §118):** llm-to-jev proposals still require the validation gate. This is a conversion assistant, not an automatic guarantee of equivalent behavior. Generated instructions and criteria must be reviewed before production use. Roadmap fixture-eval is not landed. Heuristic conversion ≠ calibrated Noul. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; 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 User-provided 0922 (`notes.md` §117 / items 330–336 / batch #100). SemIf was formerly OpenJev; MLX backend; 5.21× systems≠semantic; Softmax over options ≠ calibrated Noul; live REST 2282★. Do not reopen #23–#36/#38; do not push onto open #39/#40. 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/.agents/skills/augustus/scripts/evaluate_decisions.py b/.agents/skills/augustus/scripts/evaluate_decisions.py index e46e74b..f9210c6 100755 --- a/.agents/skills/augustus/scripts/evaluate_decisions.py +++ b/.agents/skills/augustus/scripts/evaluate_decisions.py @@ -168,15 +168,18 @@ def hysteresis_actuate(p_series, enter, exit, start=False): def hop_ece_permutation_invariant(rows, bins=10, key="p"): """Shuffle order; equal-width ECE must not move. - Hop-ECE is permutation-invariant. Clustered regime-shift errors are - invisible to it. That is a measurement fact, not a license to skip - trajectory audits (TCE / AMS live in the deferred-crispification - paper — we only lock the invariance here). + Hop-ECE is permutation-invariant. Reverse is not a sufficient + shuffle: also interleave even/odd indices. Clustered regime-shift + errors are still invisible to it. That is a measurement fact, not + a license to skip trajectory audits (TCE / AMS live in the + deferred-crispification paper — we only lock the invariance here). """ a = ece_equal_width(rows, bins=bins, key=key) reversed_rows = list(reversed(rows)) + interleaved = rows[::2] + rows[1::2] b = ece_equal_width(reversed_rows, bins=bins, key=key) - return abs(a - b) + c = ece_equal_width(interleaved, bins=bins, key=key) + return max(abs(a - b), abs(a - c)) def self_test(): diff --git a/.agents/skills/augustus/scripts/uniqueness_gate.py b/.agents/skills/augustus/scripts/uniqueness_gate.py index 2ecc213..e14d8bd 100644 --- a/.agents/skills/augustus/scripts/uniqueness_gate.py +++ b/.agents/skills/augustus/scripts/uniqueness_gate.py @@ -8,8 +8,12 @@ Also: YAML-parse SKILL.md frontmatter; notes.md owns §114, §115, §116, §117, and §118; composition items 289–316, 322–329, and 330–336 exist; -findings batches #97, #98, #99, #100, and #101 exist. Does not fetch the -network. Does not treat a lock as a Harbor score. +findings batches #97, #98, #99, #100, and #101 exist. 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. The 0843 ecosystem blurb +cites notes.md §114. Does not fetch the network. Does not treat a lock +as a Harbor score. """ from pathlib import Path @@ -48,7 +52,7 @@ "Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114" + "do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35; notes.md §114" ) UNIQ_0915 = ( @@ -249,6 +253,34 @@ def main() -> int: ): if frag not in proto_line: failed.append(f"SKILL.md protocol missing {frag!r}") + changelog = (ROOT / "CHANGELOG.md").read_text(encoding="utf-8") + for name, lock in ( + ("0843", UNIQ_0843), + ("0915", UNIQ_0915), + ("jcr", UNIQ_JCR), + ("0922", UNIQ_0922), + ("0940", UNIQ_0940), + ): + if lock in changelog: + failed.append( + f"CHANGELOG.md holds {name} uniqueness dump " + "(v0.4.0: dumps live in changelog-hourly.md)" + ) + if "Hourly 0743 uniqueness lock:" in changelog: + failed.append("CHANGELOG.md holds 0743 uniqueness dump") + readme = (ROOT / "README.md").read_text(encoding="utf-8") + if "Hourly 0743 uniqueness lock:" in readme: + failed.append("README.md holds 0743 uniqueness dump (#33 map + later lock lines)") + index = (ROOT / "docs/index.md").read_text(encoding="utf-8") + for s in ("TypeSafe Jev Choice/Score/Noul", "Install the skill"): + if s not in index: + failed.append(f"docs/index.md missing Pages gate {s!r}") + layout = (ROOT / "docs/_layouts/default.html").read_text(encoding="utf-8") + if "LICENSE" not in layout: + failed.append("docs/_layouts/default.html missing LICENSE (Pages gate after #34)") + eco = (ROOT / "docs/ecosystem.md").read_text(encoding="utf-8") + if "decision-circuits tutorial. `notes.md` §114." not in eco: + failed.append("docs/ecosystem.md 0843 blurb must cite notes.md §114") if failed: print("uniqueness-gate FAIL") for line in failed: diff --git a/CHANGELOG.md b/CHANGELOG.md index 537429b..bb808ac 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -16,85 +16,12 @@ folds: `research/notes.md`. ## [Unreleased] -User-provided 0922 HIGH (`research/notes.md` §117 / composition -items 330–336 / findings batch #100). SemIf rename + MLX + -accuracy-ladder densify onto latest main after merged **#38** -(jcr / §116). Does **not** bump the 0.4.0 pin. -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. Merged #38 owns §116 / 309–316 / #99. -Merged #35 owns §114 / 289–302 / #97. Merged #36 owns §115 / -303–308 / #98 — leave them alone. Open #40 claims §118 / 322–329 / #101. - -### Added - -- **SemIf densify (PRIMARY, `notes.md` §117).** [TheoLeeCJ/SemIf](https://github.com/TheoLeeCJ/SemIf) - MIT; homepage openjev.com; default **master**; live REST - **2282★** / **140** forks; HEAD `ca3ba65f1429` (Tolerate float - roundoff in MLX evidence verification, 2026-09-19). SemIf was - formerly OpenJev; independent; not affiliated with Jev or - TypeSafe. Interface pattern reproduction with open models; - does not reproduce Jev undisclosed model/training. Direct option - logits; 0 output tokens; shared-state parallel; MLX backend - (`--backend mlx`). Speed *theirs* 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. Browser ladder *theirs*: authored BA 0.813, - pert 0.766, TypeSafe subset 0.845 vs Published Jev 0.883 - (102 across 20 cases). Softmax over options ≠ calibrated Noul; - wire/agreement ≠ replica. SemIf ≠ kw2828/OpenJev playground ≠ - zhihz/openjev ≠ apiplant/semif-rs ≠ dddanielliu/semif-serve. - Rename is densify not a second census. JevBench 74.6 is §78 - not this ladder. Skip Archer rewrite. Quote *theirs*. - `invented_signal: false`. Composition items 330–336 / batch #100. - Uniqueness dump in [`research/changelog-hourly.md`](research/changelog-hourly.md). - - -Hourly 0843 HIGH (`research/notes.md` §114 / composition -items 289–302 / findings batch #97) plus merged #36 NanoJev -(`research/notes.md` §115 / items 303–308 / batch #98) plus -merged #38 NiazMorshed2007/jcr (`research/notes.md` -§116 / composition items 309–316 / findings batch #99) plus -user-provided HIGH alexwestco/llm-to-jev (`research/notes.md` -§118 / composition items 322–329 / findings batch #101). Does -**not** bump the 0.4.0 pin. Do not reopen or amend PR #23–#38. -Merged #37 owns §117 / 330–336 / #100. Open #39 is §114 hygiene -(do not reopen). Do not push onto open #39/#41. - -### Added - -- **NanoJev unified-games-v1 densify** — - TianyuCodings/NanoJev (Python MIT; **1289★** / **158** - forks / size **64035**; HEAD `618cea6d`; README SHA - `4190093c`). Quote *theirs*: A 0.6B parallel decision - model: states and questions in, complete probability - distributions out. Zero output-token decoding. - September 20, 2026 — One model, four games. ViZDoom - Basic 128/128 vs Jev 56/128; Predict Position 27/128 - vs Jev 11/128; Maze 225 vs Jev 2,738; Snake 30 food / - 256 steps; held-out Maze 4/10 Snake 8/8 Basic 128/128 - Predict 27/128. Untuned Qwen3-0.6B baseline. 18,760 - questions per variant; 16,333 ViZDoom; 896 Predict - Position expert episodes; hard_lr1e5; mix 1/3, 1/3, - 1/6, 1/6. Hub C-Tianyu/NanoJev revision - unified-games-v1 likes 58; dataset likes 5. not - TypeSafe Jev; open replica / specialist gameplay S1. - caijinchun/nanojev-arena ≠ liao96312/jev-arena-nanojev - ≠ zwliJay/jev-forge ≠ NanoJev. local type boolean ≠ - TypeSafe noul. Game success ≠ calibrated Noul. soft - scores ≠ hard gates. Demo HTTP 401. Quote *theirs*. - `invented_signal: false`. -- Uniqueness dump for this fold lives in - [`research/changelog-hourly.md`](research/changelog-hourly.md). - User-provided 0915 uniqueness lock: TianyuCodings/NanoJev unified-games-v1 densify; A 0.6B parallel decision model: states and questions in, complete probability distributions out. Zero output-token decoding.; One model, four games; ViZDoom Basic 128/128 vs Jev 56/128; Predict Position 27/128 vs Jev 11/128; Maze 225 attempts vs Jev 2738; Snake 30 food / 256 steps; held-out Maze 4/10 Snake 8/8 Basic 128/128 Predict 27/128; Untuned Qwen3-0.6B baseline; 18,760 questions per variant; 16,333 ViZDoom; 896 Predict Position expert episodes; hard_lr1e5; mix weights 1/3, 1/3, 1/6, 1/6; Hub C-Tianyu/NanoJev revision unified-games-v1 likes 58; dataset C-Tianyu/NanoJev-Data likes 5; HEAD 618cea6d906d54e128360786d12f703fff2b1245; 1289★ / 158 forks / size 64035; README SHA 4190093c64ee75b26e9726daa3b00cbcf6d3157a; MIT; caijinchun/nanojev-arena ≠ liao96312/jev-arena-nanojev ≠ zwliJay/jev-forge ≠ NanoJev; not TypeSafe Jev; open replica / specialist gameplay S1; soft scores ≠ hard gates; Game success ≠ calibrated Noul; local type boolean ≠ TypeSafe noul; invented_signal false; do not reopen or amend PR #31/#32/#33/#35; notes.md §115 - - -User-provided 0940 HIGH (`research/notes.md` §118 / -composition items 322–329 / findings batch #101). Does -**not** bump the 0.4.0 pin. Rebased onto merged **#37** -(SemIf / §117 / 330–336 / #100). Merged #36 owns §115 / -items 303–308 / batch #98. Merged #38 owns §116 / -309–316 / #99. Do **not** push onto open #39/#41. -Do not reopen #39. +Merged after v0.4.0: 0843 (§114 / 289–302 / #97), NanoJev (§115 / +303–308 / #98), jcr (§116 / 309–316 / #99), SemIf (§117 / 330–336 / +#100), llm-to-jev (§118 / 322–329 / #101), plus this hygiene pass. +Does **not** bump the 0.4.0 pin. Uniqueness dumps live in +[`research/changelog-hourly.md`](research/changelog-hourly.md). +Do not reopen or amend PR #23–#40. Do not push onto open #41. ### Added @@ -106,77 +33,57 @@ Do not reopen #39. convertibility; prose stays with the LLM. Heuristic conversion ≠ calibrated Noul. Companion to the decision-design card + validation gate. Composition items 322–329 / batch #101. - **HARD RULE:** do not reopen or amend PR #23–#38. Merged #36 - owns §115. Merged #38 owns §116. Merged #37 owns §117. - Do not push onto open #39/#41. Does **not** bump 0.4.0. - -Hourly 0743 HIGH (`research/notes.md` §113 / composition -items 273–288 / findings batch #96). Does **not** bump -the 0.4.0 pin. Do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#32/#33. - -### Added - -- **Hourly 0843 HIGH (`notes.md` §114).** Measurement / judgment fold - off `main` after v0.4.0. PRIMARY: pretrained Qwen2.5 base ECE already - low; instruct-tuning wrecks honesty (acc flat, mean conf 74.1%→96.7%). - Equal-width ECE ≠ quantile ECE (0.113 vs 0.076 *theirs*). Calibration - does not compose; hop-ECE is permutation-invariant (soundness theater - as a trajectory audit); Deferred Crispification; 25–60× headline - withdrawn. Hysteresis `{enter:0.8, exit:0.6}` is policy attached to a - probability, not a model property. Ranking ≠ calibration (BANKING77 - BERT-Base 93.02 is a supervised reference, not zero-shot). - g0runmezadam/what-is-jev **is** tunahansahin897/what-is-jev (rename). - Qwen2.5 / Qwen 3.8 / Qwen/Qwen3.8-27B ≠ Archer. Archer still - promised_not_landed. Evaluator: equal-width vs quantile ECE, hysteresis, - ranking≠calibration, hop-ECE invariance, cost-optimal threshold. - uniqueness_gate.py. Composition items 289–302 / batch #97. - **HARD RULE:** do not reopen or amend PR #23–#33. Merged #31 owns - `notes.md` §113 / items 273–288 / batch #96 — leave it alone. -- Uniqueness lock archive: `research/changelog-hourly.md` (this hour's - lock is there; this file stays scannable). - -- User-provided HIGH NiazMorshed2007/jcr - (`research/notes.md` §116): **Skip Archer rewrite.** - Docs-only, rebased onto merged #35 (`0189825`). - **HARD RULE:** do not reopen or amend PR #23–#35. - Merged #36 owns §115 / 303–308 / #98; open #37 owns - §117 / 315–321 / #100 (item overlap 315–316 is - #37's remap). This fold keeps §116 / items - 309–316 / batch #99. Quote *theirs*. Jev Capability - Resolver: one tool, nested capability tree, returns - context, **does not execute**. skills vs capabilities. - 0.6 band is application policy. routing ≠ permission. - docs ≠ authority to run. sol-vs-opus5-20 lookup+explain - only; n=1; wall-time mixed; Not Harbor task-execution. - Live REST: **4★**; HEAD `138b3832`; README SHA - `2a49dbc1`; size **14850**. `invented_signal: false`. - Uniqueness dump in `research/changelog-hourly.md`. +- **SemIf densify (PRIMARY, `notes.md` §117).** [TheoLeeCJ/SemIf](https://github.com/TheoLeeCJ/SemIf) + MIT; homepage openjev.com; default **master**; live REST + **2282★** / **140** forks; HEAD `ca3ba65f1429`. SemIf was + formerly OpenJev; independent; not affiliated with Jev or + TypeSafe. Direct option logits; 0 output tokens; MLX backend + (`--backend mlx`). Speed *theirs* Qwen3.5-4B 3090: direct + 1.023s vs AR JSON 5.332s (**5.21×**); argmax agree 18/21 + (systems comparison ≠ semantic equivalence). Browser ladder + authored BA 0.813, pert 0.766, TypeSafe subset 0.845 vs + Published Jev 0.883 *theirs*. Softmax over options ≠ calibrated + Noul. Rename is densify not a second census. JevBench 74.6 is + §78 not this ladder. Composition items 330–336 / batch #100. +- **NiazMorshed2007/jcr (`notes.md` §116).** Capability-tree + lookup returns context, **does not execute**. skills vs + capabilities. 0.6 band is application policy. routing ≠ + permission. docs ≠ authority to run. sol-vs-opus5-20 + lookup+explain only; n=1; wall-time mixed; Not Harbor + task-execution. Live REST **4★**; HEAD `138b3832`; size + **14850**. Composition items 309–316 / batch #99. +- **NanoJev unified-games-v1 densify (`notes.md` §115).** + TianyuCodings/NanoJev (Python MIT; **1289★** / **158** forks; + HEAD `618cea6d`). Quote *theirs*: A 0.6B parallel decision + model; zero output-token decoding; one model, four games. + Game success ≠ calibrated Noul. not TypeSafe Jev. Composition + items 303–308 / batch #98. +- **Hourly 0843 HIGH (`notes.md` §114).** Measurement / judgment + fold. PRIMARY: pretrained Qwen2.5 base ECE already low; + instruct-tuning wrecks honesty (acc flat, mean conf + 74.1%→96.7%). Equal-width ECE ≠ quantile ECE (0.113 vs 0.076 + *theirs*). Calibration does not compose; hop-ECE is + permutation-invariant (soundness theater as a trajectory + audit); Deferred Crispification; 25–60× headline withdrawn. + Hysteresis `{enter:0.8, exit:0.6}` is policy attached to a + probability. Ranking ≠ calibration. g0runmezadam/what-is-jev + **is** tunahansahin897/what-is-jev. Qwen2.5 / Qwen 3.8 / + Qwen/Qwen3.8-27B ≠ Archer. Archer still promised_not_landed. + Composition items 289–302 / batch #97. +- **Hourly 0743 HIGH (`notes.md` §113).** Harbor-jevals PRIMARY + (ywchiu/jev_benchmark). Verdict-open-jev linear ECE floor ≠ + TypeSafe replica. DecisionOps ACT / REVIEW / FALLBACK + (provider failure is **not** a policy outcome). Lock dump: + [`research/changelog-hourly.md`](research/changelog-hourly.md). ### Changed -- `evaluate_decisions.py` reports equal-width and quantile ECE, AUC, - accuracy@0.5, and a cost-optimal threshold. Self-test covers ranking≠ - calibration, hysteresis vs single 0.7, and hop-ECE permutation - invariance. - -### Added - -- **Harbor-jevals PRIMARY** — ywchiu/jev_benchmark routing - across a conversation (Gemma 4 31B 77.0% / Jev 1.13.0 - 61.4% / Laya 322M 0.0%; restriction state 95.0% against - 84.4%; 100% schema pass / six-field joint 48.8% vs 72.8% - in `results/summary.md`, *theirs*). ≠ Running-Dolphins / - Praveenrajus. Verdict-open-jev linear ECE floor ≠ TypeSafe - replica. DecisionOps ACT / REVIEW / FALLBACK (provider - failure is **not** a policy outcome). Mintzs/rlcd-lite - (Brier on stated confidence alone is a trap). - altryne/jevify ≠ Mintzs. aiwithenoch/Jev-Skill ≠ - simplosophy/jev-skill. Qwen3.8-27B ≠ Archer. Soft Noul ≠ - hard gate. Skip Archer rewrite. Quote *theirs*. - `invented_signal: false`. -- Uniqueness dump for this fold lives in - [`research/changelog-hourly.md`](research/changelog-hourly.md). - Hourly 0743 uniqueness lock: 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/#33; notes.md §113 +- `evaluate_decisions.py` hop-ECE self-test covers reverse **and** + even/odd interleave (permutation invariance is not reverse-only). +- `uniqueness_gate.py` checks Pages strings (`LICENSE` lives in + the #34 layout) and refuses CHANGELOG/README dump walls. The + 0843 uniqueness lock includes merged #34 and #35. +- `docs/ecosystem.md` 0843 blurb cites `notes.md` §114. ## [0.4.0] - 2026-09-20 diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index a279b8e..b7000c0 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -30,8 +30,10 @@ python3 .agents/skills/augustus/scripts/evaluate_decisions.py --self-test python3 .agents/skills/augustus/scripts/uniqueness_gate.py ``` -The Pages workflow must stay green (`docs/index.md` still contains the -gate strings the check greps for). +The Pages workflow must stay green. After #34 it greps `_site/index.html` +for `TypeSafe Jev Choice/Score/Noul`, `Install the skill`, and `LICENSE` +(`LICENSE` is in `docs/_layouts/default.html`; the other two stay in +`docs/index.md`). ## Fold uniqueness @@ -41,7 +43,8 @@ re-opened as "new." Before folding: - Read `research/notes.md` and the uniqueness fragments in `.agents/skills/augustus/SKILL.md` - Do not re-fold an already-landed section as a new beat -- Do not reopen or amend an in-flight fold PR (including open #31) +- Do not reopen or amend a merged fold PR (#23–#40) +- Do not push onto an in-flight fold PR (open #41) - Pre-0.4.0 uniqueness dump: `research/changelog-hourly.md` (archive, not release notes) diff --git a/README.md b/README.md index 4ead3fa..bc0e1cc 100644 --- a/README.md +++ b/README.md @@ -62,11 +62,10 @@ One line per file. The living catalog is in the reference cards and - `research/notes.md`: living hourly catalog (dense); §118 llm-to-jev conversion assistant (heuristic on-ramp, not a replica) - `research/README.md`: evidence archive index (sources, refresh log, hourly dumps) - `research/changelog-hourly.md`: hourly uniqueness dumps after v0.3.0 +- Hourly 0743 uniqueness dump: [`research/changelog-hourly.md`](research/changelog-hourly.md) (`notes.md` §113 / batch #96). +- Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 - User-provided 0915 uniqueness lock: TianyuCodings/NanoJev unified-games-v1 densify; A 0.6B parallel decision model: states and questions in, complete probability distributions out. Zero output-token decoding.; One model, four games; ViZDoom Basic 128/128 vs Jev 56/128; Predict Position 27/128 vs Jev 11/128; Maze 225 attempts vs Jev 2738; Snake 30 food / 256 steps; held-out Maze 4/10 Snake 8/8 Basic 128/128 Predict 27/128; Untuned Qwen3-0.6B baseline; 18,760 questions per variant; 16,333 ViZDoom; 896 Predict Position expert episodes; hard_lr1e5; mix weights 1/3, 1/3, 1/6, 1/6; Hub C-Tianyu/NanoJev revision unified-games-v1 likes 58; dataset C-Tianyu/NanoJev-Data likes 5; HEAD 618cea6d906d54e128360786d12f703fff2b1245; 1289★ / 158 forks / size 64035; README SHA 4190093c64ee75b26e9726daa3b00cbcf6d3157a; MIT; caijinchun/nanojev-arena ≠ liao96312/jev-arena-nanojev ≠ zwliJay/jev-forge ≠ NanoJev; not TypeSafe Jev; open replica / specialist gameplay S1; soft scores ≠ hard gates; Game success ≠ calibrated Noul; local type boolean ≠ TypeSafe noul; invented_signal false; do not reopen or amend PR #31/#32/#33/#35; notes.md §115 - - User-provided HIGH NiazMorshed2007/jcr (`research/notes.md` §116 / items 309–316 / batch #99). Capability-tree lookup; **does not execute**. Merged #35 owns §114. Merged #36 owns §115. Open #37 §117. -- Hourly 0743 uniqueness lock: 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/#33; notes.md §113 -- Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 ## Install diff --git a/docs/ecosystem.md b/docs/ecosystem.md index fa0f064..865aecc 100644 --- a/docs/ecosystem.md +++ b/docs/ecosystem.md @@ -824,7 +824,7 @@ Architecture / mental models / question-design on-ramp, not an llm-to-jev / npm ### Hourly ~08:43 Boise 2026-09-20 / 2026-09-20T14:43Z (0843 HIGH) -Architecture / mental models / Harbor-jevals / measurement honesty, not a huncho / jeval / sokudan / deferred-crispification / decision-circuits tutorial. `notes.md` §113. Skip Archer rewrite. Quote READMEs. Soft Noul ≠ hard safety. Do **not** re-fold 0806 / §112 / 0646 / §111 / 0541 / §110 / tunahan §109 census as a sibling. Fresh PR; never reopen merged #7–**#33**. Do not reopen or amend merged #31. 0★ HIGH still gets a real card. size **0** WITH CONTENTS still gets a real card. Calibration does not compose. Ranking ≠ calibration. Measurement theater ≠ Harbor. Qwen/Qwen3.8-27B ≠ Archer. Quote live REST over watch claims. `invented_signal: false`. +Architecture / mental models / Harbor-jevals / measurement honesty, not a huncho / jeval / sokudan / deferred-crispification / decision-circuits tutorial. `notes.md` §114. Skip Archer rewrite. Quote READMEs. Soft Noul ≠ hard safety. Do **not** re-fold 0806 / §112 / 0646 / §111 / 0541 / §110 / tunahan §109 census as a sibling. Fresh PR; never reopen merged #7–**#35**. Do not reopen or amend merged #31. 0★ HIGH still gets a real card. size **0** WITH CONTENTS still gets a real card. Calibration does not compose. Ranking ≠ calibration. Measurement theater ≠ Harbor. Qwen/Qwen3.8-27B ≠ Archer. Quote live REST over watch claims. `invented_signal: false`. - **edgardcham/huncho** — TypeScript MIT; **1★**; HEAD `27ab90ca`; README SHA `784fa32a`; size **319**. A hunch is a probability with a policy attached. { enter: 0.8, exit: 0.6 } is hysteresis. replay a policy change without inference. Decision models are providers, not the product. huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch. Do not copy `npm`. - **VladUZH/jev-calibration** — Python MIT; **0★**; HEAD `4c50525e`; README SHA `4e476362`; size **0** WITH CONTENTS. PRIMARY. pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%. temperature scaling still matches it in-distribution. No Jev API was called. Qwen2.5 ≠ Archer. @@ -1121,7 +1121,7 @@ class (Jev exemplar, not monopoly), mixed architecture, formal/semi-formal placement, applied placements, method substitution, composition algebra, question-design, validation, and optimizer coupling. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 ### SemIf rename + MLX + ladder (`notes.md` §117) @@ -1136,4 +1136,4 @@ Live REST **2282★**. JevBench 74.6 is §78 not this ladder. User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; 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 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/research/archive/findings.md b/research/archive/findings.md index b79b370..74121a8 100644 --- a/research/archive/findings.md +++ b/research/archive/findings.md @@ -3141,13 +3141,13 @@ placement. `invented_signal: false`. ## Batch #97 (2026-09-20 ~14:43 UTC / ~08:43 Boise) — hourly 0843 HIGH measurement / judgment -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 Note: `research/notes.md` §114. Docs + evaluator on a fresh PR off main (`cursor/hourly-0843-augustus-fold-220d`) after merged #32 (`486e93e`, v0.4.0). **HARD RULE:** do not reopen or amend -PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33. Do **not** reply -into the finished 0743 agent. Merged #31 owns `notes.md` §113 / items 273–288 / batch #96 — leave it alone. Never reopen merged #7–**#33**. Do **not** +PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35. Do **not** reply +into the finished 0743 agent. Merged #31 owns `notes.md` §113 / items 273–288 / batch #96 — leave it alone. Never reopen merged #7–**#35**. Do **not** re-fold §112 / §111 / §109 tunahan census as a sibling. Skip Archer rewrite. Quote READMEs. Soft Noul ≠ hard safety. Augustus owns placement. `invented_signal: false`. diff --git a/research/changelog-hourly.md b/research/changelog-hourly.md index 0bb609d..b7aa0bb 100644 --- a/research/changelog-hourly.md +++ b/research/changelog-hourly.md @@ -11,7 +11,7 @@ the merged **0743 HIGH** dump (PR #31 / notes.md §113 / items 273–288 - Canonical folds: [`notes.md`](notes.md) - Do not treat numbers here as new measurements. Quote *theirs*. - Soft Noul ≠ hard gate. Ranking ≠ calibration. -- Do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33. +- Do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35. --- @@ -57,7 +57,7 @@ hard safety. Hourly 0843 HIGH (`research/notes.md` §114): **Skip Archer rewrite.** Docs + evaluator on a **fresh PR off main**. **HARD RULE:** do not -reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33. Merged #31 owns `notes.md` §113 / items 273–288 / batch #96 — leave it alone. Measurement PRIMARY: instruct-tuning +reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35. Merged #31 owns `notes.md` §113 / items 273–288 / batch #96 — leave it alone. Measurement PRIMARY: instruct-tuning honesty collapse; equal-width vs quantile ECE; calibration does not compose / Deferred Crispification. Hysteresis is policy. Ranking ≠ calibration. Soundness theater: hop-ECE permutation-invariant. @@ -3758,8 +3758,8 @@ The dated passes below are how 0.1.0 was assembled. traversal as the systematic application generator; wired into SKILL.md index + toolbox sweep. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 User-provided 0920 jcr uniqueness lock: NiazMorshed2007/jcr MIT; site https://jcr.niazmorshed.dev; topics ai-agents,jev,mcp; **4★**; HEAD `138b3832`; README SHA `2a49dbc1`; LICENSE SHA `46231303`; size **14850**; Jev Capability Resolver; one tool to find documented deterministic commands in a nested capability tree; returns context; **does not execute**; skills = workflow+judgment; capabilities = individual operations; format independent of Jev; proposed open standard exploration; classify (Jev) → optional OpenAI decompose compound → beam search geometric mean of routing probs; keep up to 3 paths ≥60% of best (JCR_BAND_RATIO 0.6); ambiguity / no-match / depth-limit explicit; soft scores ≠ hard gates; 0.6 band is application policy; routing ≠ permission; docs ≠ authority to run; sol-vs-opus5-20 *theirs*: 20 scenarios × 4 variants = 80 runs; lookup+explain only, no execution; Claude Opus 5: agent input 108,585→15,819 (−85%), cost $0.3700→$0.1222 (−67%), wall 105.5s→77.7s; Codex GPT-5.6-Sol: 61,952→47,669 (−23%), $0.1377→$0.1151 (−16%), wall 25.3s→62.4s (Sol slower with JCR in 19/20); One Sol outlier 372.6s / 193 Jev calls; n=1 per cell; Not Harbor task-execution; Claude/Codex harnesses; compare mode; 50 scenarios bundled; 11 groups, 960 nodes, 11,360 items; 16 routing rounds per step; 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 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave) diff --git a/research/notes.md b/research/notes.md index 1bb8b40..577ad8d 100644 --- a/research/notes.md +++ b/research/notes.md @@ -28088,13 +28088,13 @@ Measurement / judgment / skills fold on a **fresh PR off main** (`cursor/hourly-0843-augustus-fold-220d`) rebased onto latest `main` after merged #31 (0743, `notes.md` §113 / items 273–288 / batch #96), #32 (`486e93e`, Release v0.4.0), -and #33 (README map). **HARD RULE:** -do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33. +#33 (README map), and #34 (Pages layout). **HARD RULE:** +do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35. Do **not** reply into the finished 0743 agent. Do **not** reopen or amend merged #31. This fold's IDs: `notes.md` §114 / composition 289–302 / findings batch #97. -Never reopen merged #7–**#33**. Do **not** re-fold +Never reopen merged #7–**#35**. Do **not** re-fold §112 Merve / §111 0646 / §110 0541 / §109 0439 / §85 Akshay / §78 JevBench v1.2 board / §71 claim-audit / §60 six-gates / §109 tunahansahin897/what-is-jev *as a second census* (this @@ -28119,7 +28119,7 @@ Archer still **promised_not_landed**. Unique consecutive fragments (this hour) must appear as **one substring** in overlays (see uniqueness gate): -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 ### How-to-apply (five placements / measurement lenses) diff --git a/research/refresh-log.md b/research/refresh-log.md index d77047d..82228ef 100644 --- a/research/refresh-log.md +++ b/research/refresh-log.md @@ -3114,9 +3114,9 @@ User-provided 0922 uniqueness lock: SemIf was formerly OpenJev; independent; not ## 2026-09-20 ~14:43 UTC — hourly 0843 HIGH measurement / judgment - Docs + evaluator on a **fresh PR off main** (`cursor/hourly-0843-augustus-fold-220d`). Rebased onto latest `main` after merged #31 (0743, `notes.md` §113 / - items 273–288 / batch #96), #32 (`486e93e`, Release v0.4.0), and #33 - (README map). **HARD RULE:** do not reopen or amend PR - #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33. + items 273–288 / batch #96), #32 (`486e93e`, Release v0.4.0), #33 + (README map), and #34 (Pages layout). **HARD RULE:** do not reopen or amend PR + #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35. Do **not** reply into the finished 0743 agent. Do **not** reopen or amend merged #31. Merged #31 owns `notes.md` §113 / items 273–288 / batch #96 — leave it alone. This fold: `notes.md` §114 / items 289–302 / @@ -3137,12 +3137,12 @@ User-provided 0922 uniqueness lock: SemIf was formerly OpenJev; independent; not applied-mappings, validation, judgment-class, composition-algebra items 289–302, toolbox-mapping, methods-catalog, formal-methods, formal-semi-formal, question-design, agent-self-assessment, mappings, CHANGELOG Unreleased, changelog-hourly, README, - docs/ecosystem, findings batch #96, sources.json. Hunches labeled. No wrapper. + docs/ecosystem, findings batch #97, sources.json. Hunches labeled. No wrapper. - Quote *theirs*. Do not invent accuracy numbers. `invented_signal: false`. Do **not** merge from this fold agent. -Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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; notes.md §114 +Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) 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/#34/#35; notes.md §114 ## 2026-09-20 ~15:15 UTC — user-provided HIGH NanoJev unified-games-v1 densify - Docs-only on a **fresh PR off main** (`cursor/nanojev-unified-games-v1-densify-9049`). @@ -3187,4 +3187,4 @@ Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { across 21 overlays, uniqueness_gate.py (0843 + 0915 + jcr + 0922). Not a first sighting. No wrapper. Quote *theirs*. `invented_signal: false`. X MCP not used; no invented tweets. - 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 +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>>>>>>> 4fcac46 (Post-#35 hygiene: §114 citation, scannable dumps, lock #34/#35, hop-ECE interleave)