[24601]

Docs-only fold of hourly System One / categorization watch into Augustus how-to-apply, mental models, mixed architecture, and Harbor/jevals practice. Not a hit list. Not a stub.

Archer Hume 27B drop is still WATCH. Omni perception→decision does not wait for it.

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 test/PR/HA write/authorship on a Noul is soundness theater. Soft Noul ≠ hard safety. No wrapper, no SDK/plugin how-to, no copied ports/key-file paths. No invented metrics. TypeSafe Jev is the exemplar, not the monopoly. Causes below are hunches, not promoted Contract.

Lane is Augustus only (mental models / architecture / Harbor). 0★ HIGHs still get real cards because they were named HIGH.

This update — 17:25 UTC / ~11:25 Boise 2026-09-19 (research/notes.md §88)

Queued user SIGNALs plus the remaining deferred 0945 HIGH. laya-vision and Cerebellum stay §87 — not re-carded. Four how-to-apply clusters:

A. Open LoRA replica vs encoder namesake

1. GestaltLabs/Jeff-1 (Hub Apache-2.0; 4 likes) + Gestalt-Lab/jeff (Apache-2.0; 0★; README SHA bbf66409) — Qwen3-4B LoRA on Qwen3-4B-Instruct-2507; Choice / Noul / Score; optional POST /v1/systemone. README theirs: “API compatibility does not imply identical judgments or performance.” Confidence = max label p.

   Fact-check bake-off theirs on the same 9,730 FEVER / VitaminC / SciFact / Climate-FEVER items vs live Jev 1.13.0 (max-class-p; Jev's separate API confidence not used): Jeff acc 0.8183 ECE 0.0807 vs Jev 0.8283 / 0.0932. Acc/Brier lose; ECE wins. Set reused for error analysis — not an untouched holdout. Weak on not_enough_info. Lower ECE ≠ individual correct.

   ≠ logan-markewich/jeff GLiFormer encoder. Hub GestaltLabs ≠ GitHub Gestalt-Lab ≠ logan-markewich.

B. Jev-first bounded agent

2. devagrawal09/stanley-code (MIT; 20★; README SHA 59da9a1f) — NL request → one bounded workflow; deterministic gather + fixed-choice Jev; code owns decisions. README theirs: empty findings is not an approval; read notChecked; there is no pass/approved. Router 0.6/0.55/0.15 still soft. Human --promote-candidate only. Pi fallback unverified. jev-code 0.0.1 does nothing; 0.1.0 not on npm.

C. NL memory → beam-search FS

3. marc2332/findme (Rust; license null; 4★; README SHA f2a2ca71) — tagline theirs: “jev + fs = findme.” Beam over listed names+metadata; parent fallback ≤4; gitignore / symlink skip in code. Life/knowledge retrieval, not only SWE. ≠ Peu77/JevFind path-then-window. Ranking ≠ identity.

D. Price workers, not the conversation (remaining deferred 0945 HIGH)

4. leftspace89/jevsubrouter (MIT; 4★; README SHA 317996d0) — sub-agent dispatch binding vs turn-level advice. README theirs: “the conversation you are actually having never changes model — so its prompt cache is never invalidated.” Fail-open. Low conf → balanced, never silent down. Stats are counts, not dollars — worker tokens are invisible from a hook. ≠ jev-gateway ≠ slo-router.

Deferred 0945 already in §87 (pointer only)

• thaitea/laya-vision-smolvlm-256m — SmolVLM; score untrained; ≠ blackwood ≠ Archer
• mkeco/Cerebellum-2B — /v1/decide ≠ TypeSafe; wire-compat vs agent-routing as separate Harbor axes; competing NAR not endorsement

Name locks / anti-patterns this hour

• GestaltLabs/Jeff-1 ≠ logan-markewich/jeff; Hub GestaltLabs ≠ GitHub Gestalt-Lab
• Do not quote Jeff-1 ECE as “better than Jev” (acc/Brier lose; set reused)
• Empty stanley findings ≠ approval; do not auto-promote agent-written workflows
• findme ≠ JevFind; beam score ≠ file-identity proof
• Do not swap the conversation model to save worker $; do not quote unmeasured dollars
• Soft Noul ≠ hard safety: Jeff-1 ECE, stanley 0.6/0.55/0.15, findme beam, jevsubrouter balanced-on-low-conf are sensors
• Do not copy uv / PEFT / port 8079 / npm ci / cargo install / marketplace / TYPESAFE_API_KEY / JEVSUB_API_KEY

Reviewer entry points (§88)

1. research/notes.md §88 — full receipts (hunches labeled); quote README/Hub
2. references/mental-models.md — Apply queued 1047 follow-ons four clusters
3. references/mixed-architecture.md — fail table + gallery rows
4. references/faq.md — four new Qs
5. references/judgment-class.md — Jeff-1 after logan-markewich/jeff; when-to-use table namesake lock
6. references/validation.md — n=9730 / stanley coverage / jevsubrouter counts
7. references/formal-methods.md — follow-on envelopes outside the fill-in template (stanley-code \notChecked``)
8. SKILL.md YAML triggers + protocol + mapping-index unique fragments

Prior fold — 16:52 UTC / ~10:47 Boise (research/notes.md §87)

Thirteen 1047 HIGH signals plus three deferred 0945 class notes (decision-validated UI; decision-as-assert; hybrid S1; pointer search; Harbor trio; authorship; non-SWE HA/n8n; compaction-pi namesake lock; jevloop; laya-vision / Cerebellum / laya-grounded). See earlier commits on this branch.

Sections changed (cumulative)

• Skill: SKILL.md; mixed-architecture.md; faq.md; mental-models.md; judgment-class.md; applied-mappings.md; mappings.md; validation.md; composition-algebra.md items 27–40; question-design.md; methods-catalog.md; toolbox-mapping.md; formal-methods.md; agent-self-assessment.md; optimizer-integration.md
• Research: notes.md §87–§88; sources.json (490 sources, 487 unique URLs, retrieved 2026-09-19T17:25Z); refresh-log.md; archive/findings.md batches Fold hourly 1019 HIGH: GEPA ADE + novel cards + REVISIT densify #70–Release v0.5.1 #71
• Index: CHANGELOG.md; README.md; docs/ecosystem.md
• Unchanged: evaluate_decisions.py

Verification (HEAD 3f4ae90)

• Docs-only; no app UI to browser-test
• evaluate_decisions.py --self-test → self-test ok
• sources.json parses (490 / 487 unique; three pre-existing duplicate URLs left untouched)
• All four queued URLs + jevsubrouter present in §88
• SKILL.md YAML frontmatter parses
• git diff --check clean
• Mapping-index unique §88 fragments land consecutive in named Detail files (when-to-use table + formal-methods envelope)

Adversarial review

Pass 1 (94bd5bd): 0 Critical, 6 Important — mapping-index Detail files did not contain cards SKILL.md claimed.

Pass 2 (db5b794): those six closed; no new Critical/Important.

Pass 3 (23432f0): §88 fold; 0 Critical, 4 Important — open-weights fragment not in when-to-use table; stanley-code \notChecked`` missing from formal-methods; findme cited item 27 instead of JevFind item 30; jevsubrouter README quote was not verbatim.

Pass 4 (3f4ae90): those four closed; no new Critical/Important; no merge-blocker.

Harbor honesty held: Jeff-1 acc/Brier lose ECE wins on a reused set; stanley empty ≠ approve; findme ranking ≠ identity; jevsubrouter counts ≠ dollars; Soft Noul ≠ hard safety.

  

[24601]

Adversarial review — PASS (nits only; not merge-blockers)

Read-only pass on 3f4ae90 vs main. Spot-checked cited numbers against live READMEs/Hub cards (GestaltLabs/Jeff-1, Gestalt-Lab/jeff, stanley-code, findme, jevsubrouter, jev-frontier-bench, jev-gliclass-bench, job-posting-triage, laya-vision, Cerebellum README_EN, laya-grounded). Did not merge.

Checklist

#	Gate	Result
1	Invented metrics / unattributed numbers	PASS. Jeff-1 n=9730 acc 0.8183 / ECE 0.0807 vs Jev 0.8283 / 0.0932 matches Hub+README (max-class-p; set reused). Frontier-bench 72.5%/0.161 vs Fable 84.0%/0.064, cascade 82.5% at $4.41/1k in-sample, ChaosNLI JS 0.149 vs uniform 0.127 match. GLiClass 78/40/49 n=100 match. Job-posting floor 0.947 / tfidf 0.970 F1 0.700 / Jev ECE 0.046 / llm_local 1000/1000 conf 1.0 ECE 0.947 match. Laya-vision val n=8235 acc 75.2% ECE cal 0.034; score untrained; SigLIP projector “never learned to use the image” is in r33drichards/laya-vision README. Cerebellum 94.92% vs Jev 81.1% quoted theirs unverified. Laya-grounded phishing 0.611→0.512 / Platt not temperature match. Stanley 0.6/0.55/0.15 is the README mermaid gate. Findme parent fallback ≤4 match. Jevsubrouter ~300 ms warm / counts ≠ dollars match.
2	Class-table inflation (multimodal/omni on text-only)	PASS. Jeff-1 / stanley / findme / jevsubrouter stay text. Anima3 is a11y tree, not a screenshot. Laya-vision is actually SmolVLM and is locked ≠ blackwood ≠ Archer, not dropped into the Archer/omni row. Cerebellum / laya-grounded called text-state NAR. Skip-Archer block at research/notes.md:13346.
3	Name collisions	PASS. GestaltLabs/Jeff-1 ≠ logan-markewich/jeff (GLiFormer) ≠ Gestalt-Lab/jeff GitHub. Findme ≠ JevFind (item 30). ha-switchboard ≠ HA-Jev. fast-jev-compaction-pi ≠ pi-jev-compact ≠ pi-jev-compaction. mkeco GitHub ≠ mkzero Hub. choxos ≠ egma left intact. jev-frontier-bench ≠ jev-frontier-100.
4	Archer Hume false positives	PASS. Watch throughout. Qwen3-4B Jeff-1 is not the 27B drop. Laya-vision SmolVLM is not Archer. Cerebellum Qwen3.5-2B is not Archer.
5	Soft Noul / Choice as hard gate without envelope	PASS. Jevtest 0.85 / anima3 0.35 / n8n 0.5 / JevFind 0.25/0.55 / stanley 0.6/0.55/0.15 / Cerebellum ActEscalate 0.50 labeled still soft. Envelopes named (Telegram 4096 / callback_data; HA allowlist/freshness/idempotency/post-state; toContain; closed verb menus; gitignore/symlink skip; prompt-cache never swap). Hard-gating labeled soundness theater.
6	Silent fallback / soundness theater	PASS. Jevsubrouter: low conf → balanced, never silent down; fail-open. Compaction-pi fail-open to built-in LLM summary (named). n8n Low Confidence is a separate output, not silent-best-route. Empty stanley findings ≠ approval.
7	Duplicate sections already on main	PASS. git grep on origin/main has no gram-render / Jeff-1 / jevsubrouter / laya-vision-smolvlm cards. fast-jev-compaction-pi was a MED one-liner; this fold is the productization (same move as jev-semgrep), not a re-card of a §87-on-main section.
8	Contradictions within the PR	PASS. §87 vs §88 agree (laya-vision + Cerebellum stay §87, not re-carded; jevsubrouter is the remaining 0945 HIGH). Acc/Brier lose + ECE wins is consistent; never quoted as “better than Jev.” Anima3 README “No economy” vs Economy section is labeled theirs.

Harbor honesty held: Jeff-1 reused set; stanley empty ≠ approve; findme ranking ≠ identity; jevsubrouter counts ≠ dollars; Soft Noul ≠ hard safety.

Nits (do not block merge)

1. Mapping-index Detail vs landing table. SKILL.md open-weights row cites references/judgment-class.md (when-to-use table) for the unique fragment ≠ GestaltLabs/Jeff-1 Qwen3-4B LoRA replica. That exact lock is in the families table Watch column (judgment-class.md:41), not the when-to-use grid (judgment-class.md:585). The full Jeff-1 card is later at judgment-class.md:880. Pass 4 claimed a when-to-use retarget; the fragment still sits on the GLiFormer families row. Content is correct; the Detail pointer is stale.

2. Species map still unqualified jeff. judgment-class.md:59 still says jeff (GLiFormer) after introducing GestaltLabs/Jeff-1. Cards and FAQ lock the names; the species map does not. Not a collapse — an incomplete lock.

3. Jevsubrouter verbatim vs paraphrase. README theirs: “the conversation you are actually having never changes model — so its prompt cache is never invalidated.” That sentence is quoted in research/notes.md:13603. FAQ (faq.md:2215) and the mixed-architecture fail row (mixed-architecture.md:232) paraphrase. Fine for how-to-apply; not the verbatim landing Pass 3 asked for in every surface.

Parent can merge. No further commits from this pass.