Fold hourly 0843 HIGH: hysteresis, instruct-tuning ECE, calibration does not compose

.agents/skills/augustus/references/agent-self-assessment.md

@@ -924,5 +924,9 @@ is Watch / empty repo this pass (`notes.md` §51).
   ALLOW ([interlock](https://github.com/somoore/interlock) *kernel* —
   Jev is SENSOR, policy decides; `notes.md` §59), ok / retry /
   escalate / stop (harnessjudge). Same shape as the lifecycle gates above.
+
+Hourly 0843 agent self-assessment: omni-/ask-jev probabilities are advisory, not calibrated guarantees. 13 focused checks and one mutually exclusive outcome. Do not treat low mutation-attribution p as identifying someone else. huncho journal/replay: change policy without re-inference. `notes.md` §114.
 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

.agents/skills/augustus/references/applied-mappings.md

@@ -2470,4 +2470,23 @@ Hourly 0646 uniqueness lock: Calibration is not alpha; NO CURRENT ALPHA CANDIDAT
 **User-provided 0806 HIGH (`notes.md` §112).** Institutional HF voice / encoder ZS lineage:
 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
 people who compare Jev against GPT-5.6 has never fine-tuned BERTForXYZ for living and it shows. Jev vs GPT-5.6 bakeoffs are a category error. encoder / ZS classifiers. opt for DeBERTa and ModernBERT ones. multimodal image<>text ZS as perception front-end. softmax/ZS scores still ≠ calibrated Noul. soft scores ≠ hard gates. Do not copy `pipeline()` / `pip`.
+
+**Hourly 0843 HIGH (`notes.md` §114).** Measurement PRIMARY + hysteresis/policy + catalogs-as-lenses:
+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; ranking ≠ calibration.
+calibration does not compose; Deferred Crispification; TCE / AMS; 25–60× headline withdrawn.
+pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet.
+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; catalog ≠ endorsement.
+omni-/ask-jev ≠ pedroknigge/mcp_jev; Probabilities are advisory, not calibrated guarantees.
+light_cutoff_applied_to_combination 0; recorded run, kinematic animation; Bring your own API key.
+AND: product (independence assumed and recorded in the trace); circuit-vl-4b ≠ Archer.
+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 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