[24601]

Docs-only fold of hourly System One / categorization watch 1347 (Boise ~13:47 MDT / 19:47 UTC) into Augustus how-to-apply, mental models, mixed architecture, toolbelt, and Harbor/jevals practice. Not a thin Jev skill dump. Not a hit list.

PR #3 / #4 / #5 / #7 already merged. This is a fresh PR off main. Never reopen merged PRs. Do not re-fold 1241 / §90.

Archer Hume 27B drop is still WATCH (HF empty; tracker likes 49 lastModified 2026-09-19T18:37:18Z still promised). Laya yes. Blackwood ABSENT. Census user-provided, not re-derived: SemIf 1846 (+17), jevlike 962 (+3), TypeAR-AI/TypeAR 10 (+1), Awesomejev flat 561/27007.

Non-negotiable: exact work stays in code or policy; the model owns narrow judgment; never launder a soft Noul as a proof. Formal methods compose with categorization/scoring; a Noul is a SENSOR. Hard-gating argmax as “safety”, pasting 97.7% n=130 / 0 hallucination / 8,026 tokens as class ceilings, treating a skill named System One as a judge, or collapsing hermes-switchyard into Agnes/WHETHER-HOW-WHAT is soundness theater. Soft Noul ≠ hard safety. No wrapper, no SDK/plugin how-to, no copied TYPESAFE_API_KEY / OPENROUTER_API_KEY / JEV_API_KEY / uv / npx / plugin-marketplace install / .env. No invented metrics. TypeSafe Jev is the exemplar, not the monopoly. Backend-agnostic categorization/scoring/decision class (Jev-like speed/econ).

Lane is Augustus only (mental models / architecture / Harbor-jevals / toolbelt). 0★ HIGHs still get real cards because they were named HIGH. FrancoisChastel/jev-code is the only 1★ this hour.

Adversarial review on this PR: FAIL on 9a8f46d, fixed on 3d89dfa (closed-vote stuffing, generate.py vs systemone generate, live pi-jev-gate polarity). Parent merges.

This update — 19:47 UTC / ~13:47 Boise 2026-09-19 (research/notes.md §91)

Seventeen HIGH repos plus two rh-guard cousins (cross-ref only).

1. Judge harness as control API

lexingtonhibiki/judgekit (MIT; 0★; README SHA 0fcb42ea) — YAML task → classify/score/route/verify. Native TypeSafe or OpenAI-compat or keyword rules. n=130 zh mini sets theirs: Jev 97.7% [93.4–99.2%] (127/130; 3 runs, 0 flips) ~890 ms / ¥0.105/1k. 0.7 gate caught 100% of 3 errors at ~9% escalation. ≠ JudgeBench ≠ DeepEval ≠ promptfoo.

Ascurse/typed-judge-kit (MIT; 0★; README SHA 2bfa614d) — typed questions → one batched call → verdict in your combine(). Thresholds from labels, not self-confidence. 7.0× (1.56 s vs 10.92 s); MIN_LABELS=20 / P_MIN=0.95. Noul rerank made R@5 worse 14/15→13/15 theirs. ≠ judgekit.

2. Batch packing VOI

alsoleg89/decide (license null; 0★; README SHA fd8aafcf) — MCP bulk classify; compact summary + review queue. 500 VS Code issues theirs: $0.0203 / 39.5 s; JSON −96.5% / −49.3%; 204 review @0.8. 0.8 ≠ 80% accuracy. ≠ jev-sift.

3. Calibration as product

AnthusAI/Jev-Calibration (license null; 0★; README SHA bc86c044) — raw ECE 0.117 → Platt 0.052 → isotonic 0.008 theirs. Choice 50–95% stated conf sits at ~50–57% accuracy. confidence field ≠ top-p.

pmcclelland/jev-calibration-arena (license null; 0★; README SHA 26498a10; size 0) — Brier + reliability. never acts. Scaffold blocked. ≠ jev-arena ≠ jevarena.

4. Decision-as-Plugin

ctmx/openrouter-jev-mcp (MIT; 0★) — OpenRouter alpha Decisions. Native TypeSafe keys not supported. A judgement does not grant permission to act.

cyrusasco/typesafe-mcp (MIT; 0★) — ts_decide ≥0.75 auto; noul deadband 0.35–0.65. ts_safety is deterministic patterns.json.

FrancoisChastel/jev-code (MIT; 1★) — ≠ npm jev-code 0.0.1 (stanley placeholder). auto_accept 0.85 still soft.

skylence-org/claudecode-jev-marketplace (license null; 0★; size 0) — event-boundary only; fail-open not hot path.

pedroknigge/mcp_jev (MIT; 0★) — packs not ask_jev. Key once in ~/.mcp_jev.

codaaiteam/jev-skill (license null; 0★) — independent /api/v1/decide at jevtypesafeai.com. ≠ TypeSafe.

5. Policy-constrained skill select

bgrablin/hermes-switchyard (MIT; 0★; v0.4.1) — advisory; never loads skills; never changes the model. ≠ hermes-jev-router ≠ hermes-plugin-jev (Agnes).

6. Tiny local econ pruner

dmdjr1409/nanoprune (MIT; 0★) — 2.8MB / 495k params / ECE 2.58% theirs. Distilled from Laya 421M. “0 hallucination guaranteed” is soundness theater.

7. Observe→score→act cousins

smartdio/jev-browser-agent (MIT; 0★) — ≠ ZHUBoer/ego-jev. conf<0.6 escalates. DONE from evidence. Not closed-vote (outer LLM still plans).

Dakai/omp-jev-web (MIT; 0★) — overlay 5.9 s / 0 refusals vs upstream 47.6 s / 27 theirs. DONE ≠ proof. ≠ omp-greenlight. OMP agent still plans (not closed-vote).

hari007sh/jev (license null; 0★) — local cross-encoder + voice CU. ≠ dannote/jev. Teacher is systemone generate. Local encoder still plans (not closed-vote).

8. Deterministic verify ≠ System One

0thernet/system-one-skills (MIT; 0★) — only system-one-verify ships; no model call. 8,026 tok theirs on 3/24. Name ≠ species.

9. Soft-score vs hard-argmax (rh-guard cousins; cross-ref)

harshpuri84/typed-gate (MIT; 0★; size 0) — band [0.40,0.60] is refusal. 0.51 is not a yes. 0 wrong / 0 omit / 117 review theirs. rh-guard owns the gate cousin.

fivethirty/pi-jev-gate (MIT; 0★; size 0; HEAD 902f4e6; README SHA 62482af1) — fail-closed; choice is the verdict. No local threshold. Checker-down / missing credential / no verdict blocks. Ask band removed. ≠ pi-jev-approver (also fail-closed; different product). Still a sensor; rh-guard owns.

Files

SKILL.md mapping index unique fragments; mental-models.md Apply 1347; mixed-architecture.md fail table + gallery; faq.md; judgment-class.md; validation.md; applied-mappings.md; mappings.md; toolbox-mapping.md; composition-algebra.md items 62–70; question-design.md; methods-catalog.md; formal-methods.md; agent-self-assessment.md; CHANGELOG / README / docs/ecosystem.md; research/notes.md §91; refresh-log.md; archive/findings.md batch #74; sources.json (543 / 540 unique, retrieved 2026-09-19T19:47Z).

Checks

• evaluate_decisions.py --self-test ok
• SKILL.md YAML frontmatter parses
• Unique mapping-index fragments consecutive in SKILL.md + notes.md §91 + composition-algebra.md + mental-models.md + faq.md (including pi-jev-gate fail-closed; choice is the verdict)
• sources.json 543 sources / 540 unique (3 pre-existing dups unchanged)
• git diff --check clean
• Did not invent stars or Harbor claims. Did not reopen merged Fold hourly 1047 HIGH + queued SIGNALs + deferred 0945 into Augustus #3–Fold hourly 1241 HIGH into Augustus mental models and jevals #7. Did not re-fold 1241 / §90. Did not fold jev-omni HIGH or extra rh-guard HIGH this pass.

  

[cursor]

Adversarial review — FAIL on 9a8f46d, fixed on 3d89dfa (not merge-blockers now)

Read-only first on gh pr view 8 / gh pr diff 8 vs origin/main (after merged #7). Actual fold is seventeen 1347 HIGH receipts + two rh-guard cousins into research/notes.md §91 and skill docs (mental models / mixed architecture / Harbor-jevals) — not a thin Jev skill dump. Spot-checked cited numbers against pinned READMEs at the SHAs in research/notes.md §91. Live fivethirty/pi-jev-gate HEAD 902f4e603efea759fcc22c36ba3beba600dad7ab README SHA 62482af1 is fail-closed; choice === "block"|"allow" is the verdict; no local threshold in extensions/index.ts; checker-down / missing credential / no verdict blocks. hari007sh/jev teacher is systemone generate (no generate.py in the tree). evaluate_decisions.py --self-test → self-test ok. YAML frontmatter parses. Unique fragments consecutive in SKILL.md + notes.md §91 + composition-algebra.md + mental-models.md + faq.md. Did not merge.

Checklist

#	Gate	Result
1	Invented metrics / unattributed numbers	PASS. theirs numbers match pins: judgekit 97.7% [93.4–99.2%] n=130 / ~890 ms / ¥0.105/1k; typed-judge-kit 7.0× (1.56 s vs 10.92 s) / MIN_LABELS=20; decide $0.0203 / 39.5 s / JSON −96.5% / −49.3% / 204@0.8; Jev-Calibration ECE 0.117→0.052→0.008; typed-gate 0 wrong / 0 omit / 117 review vs argmax 25 omit / Haiku 6 wrong; omp-jev-web overlay 5.9 s / 0 refusals vs 47.6 s / 27; nanoprune 2.8MB ECE 2.58%; system-one-skills 8,026 tok on 3/24 (holdout would add 3,612). 0.8 ≠ 80% accuracy is labeled. Noul AUROC 0.83 is README rounding of 0.828 (nit).
2	Class-table inflation	PASS. references/judgment-class.md:986 §91 bullet is “placements / unofficial packaging, not new species”. No new omni/Archer/multimodal species. nanoprune is Laya 421M distill, not hosted Jev. system-one-skills is deterministic verify, not a judge.
3	Name collisions	FAIL → fixed. Closed-vote cell stuffed observe→score→act cousins (same class as PR #7 ZHUBoer FAIL). smartdio/jev-browser-agent escalates to the outer LLM; Dakai/omp-jev-web has an OMP agent; hari007sh local encoder still plans — none is “no planner LLM”. Removed from SKILL.md:133 Closed-vote; they stay on Search/control SKILL.md:146. applied-mappings.md:1572-1585 now says not closed-vote (outer LLM / OMP agent / local encoder still plans). Other locks hold: FrancoisChastel/jev-code ≠ npm jev-code; hermes-switchyard ≠ hermes-jev-router ≠ hermes-plugin-jev; smartdio/jev-browser-agent ≠ ZHUBoer/ego-jev; hari007sh/jev ≠ dannote/jev; jevtypesafeai.com ≠ TypeSafe. Hot-click stays jiangkoumo-only (methods-catalog.md:125).
4	Archer Hume false positives	PASS. Skip Archer / still NOT landed (research/notes.md:15095-15098; census block). nanoprune is Laya 421M, not the 27B drop. None of the seventeen is treated as Archer.
5	Soft Noul as hard gate without envelope	PASS after polarity fix. Thresholds labeled sensors (research/notes.md:15082-15087; formal-methods.md:688-693). typed-gate band [0.40,0.60] is refusal. Live pi-jev-gate has no local 0.50 threshold; Choice is the verdict behind a read-only allowlist, fail-closed on errors. Still a sensor; rh-guard owns (notes.md:15046-15068). Soft Noul ≠ hard safety: choice-as-verdict without a mid-band is still soft judgment acting as a hard stop — Augustus cross-refs, does not copy pi install.
6	Silent fallback / soundness theater	PASS. Hard-gating argmax as safety / pasting 97.7% n=130 / 0 hallucination / 8,026 tokens as class ceilings / treating a skill named System One as a judge / collapsing hermes-switchyard into Agnes is labeled soundness theater (notes.md:15087-15091). Decision-as-Plugin fail-open missing key / not-hot-path. Arena never acts. claudecode-jev-marketplace fail-open not hot path.
7	Duplicate sections already on main after #7	PASS. Fresh §91. Do not re-fold 1241 / §90. Do not re-card ZHUBoer/ego-jev / jev-ultrafast / omp-greenlight / hermes-jev-router / hermes-plugin-jev / jev-sift / stanley jev-code 0.0.1 / dannote/jev.
8	Contradictions within the PR	FAIL → fixed. (a) Teacher: faq/mixed-arch said generate.py; notes + hari007sh README say systemone generate (no generate.py in tree). Now faq.md:2491-2492, mixed-architecture.md:277, applied-mappings.md:1585, notes.md:15000. (b) Polarity: faq asked “Is pi-jev-gate fail-closed?” and answered “No, and no” while the unique fragment and live HEAD are fail-closed; notes still said “Fail-open if the checker errors” and p(block)≥0.50. Now heading is “Is pi-jev-gate fail-open?” (faq.md:2504-2506); fail-table / question-design treat fail-open as the anti-pattern (mixed-architecture.md:280, question-design.md:195); pin HEAD 902f4e6 README SHA 62482af1 (notes.md:15048, sources.json:3261, docs/ecosystem.md:598). Unique fragment pi-jev-gate fail-closed; choice is the verdict is consecutive in the five required files. ≠ pi-jev-approver kept (also fail-closed; different product).

Nits (not blockers)

• Noul AUROC 0.83 vs Llama 0.72 (notes.md:14844-14845) is README rounding of 0.828 / 0.719. Both Noul and Choice AUROC are 0.828 on the pin; do not treat 0.83 as a second measurement.
• judgekit 0.7 gate that “caught 100% of 3 errors” is the keyword-rule fallback arm, not hosted Jev. Still labeled a sensor.
• hermes-switchyard: transport-failure may preserve the local winner (advisory; never loads skills). Hosted routing not claimed.
• alsoleg89/decide pin-vs-live SHA drift; $0.0203 / 39.5 s / 96.5% / 49.3% / 204@0.8 still match the pin.
• jev-calibration-arena pin README is a stub; SPEC.md has college_football / ~0.6 / kill-100. never acts still holds. ≠ jev-arena.
• Live pi-jev-gate README leftover “block threshold are hardcoded constants” vs extensions/index.ts (no local threshold; choice is the verdict). Fold followed the code + “Fail-closed” / “decision is final” README claims.
• Closed-vote mapping cell still says “hot-click cousin: ego-jev” without the jiangkoumo owner (pre-existing on main; Fold hourly 1241 HIGH into Augustus mental models and jevals #7 nit). 1347 cousins are not in that cell.
• FrancoisChastel/jev-code is the only 1★ this hour; 0★ HIGHs still get real cards.

Parent merges. Ready for parent merge on 3d89dfa.