Fold hourly 1203 HIGH: Swift 6 Foundation Models bridge + novel cards + REVISIT densify

[24601]

Summary

Hourly 1203 fold, rebased onto post-#73 main (8a096a8). Glance owns notes §147. This fold is notes §148, composition 761-776, findings batch #129. Does not re-card glance. Does not bump 0.5.1. Does not reopen #23-#73. Leave this draft. Parent merges only after ADV_PASS.

TypeSafe Jev stays the dominant exemplar and the default recommended path. The lane is still the broader decision-model class: classifiers, encoders and decoders, specialized AR and constrained heads, vision and listwise scorers, and what TypeSafe calls System One. Soft judgment never sole veto.

PRIMARY: peterfriese/jev-foundation-models (1 star, Apache-2.0, HEAD 27963995965d). Swift 6 bridge for TypeSafe Jev System One into Apple Foundation Models. README claims of 40-150 ms, zero hallucinations, and calibrated probabilities are theirs, not rerun. Never embed API keys. Bool maps to noul, enum to choice, range guide to score. A bridge is not an on-device replica. Jev remains the decision model.

Also carded: Alexander-Ollman/laya-ft (Aegis / ToxicChat / WildGuard / BeaverTails F1 and XSTest false alarms 24.8% to 65.2% to 47.2% are theirs; detection gain is not a license to auto-block), Codercise/jev-in-practice (key stays in Node; playground is not a bench; no accuracy table invented), Ejokey/lightjev (one crawl, 6 pages, 4 hits, $0.000851 theirs; one crawl is not Harbor), Renwang-Huang/arbitype (independent, not official TypeSafe; soft gate is not a sole veto), ruslanlap/jev-gate (sum to 1 within 0.02 else exit 2; noul p=0.02 is not a merge), emlama/jev-mcp (saved tool is inputs plus context plus questions plus docs; one SQLite volume).

REVISIT densify, marked theirs, prior section kept:

• AkashPriyadarshii/jev-seo §131 (HEAD f42455ac951a; catalog is not endorsement; star move is not the fold)
• dtduc-git/jevnav §129 (HEAD b7a12d2f54ce; page truth, not pixels; star 0 to 1 is star-noise)
• SoundBlaster/SwiftDecision first card (revisit tag had no prior notes card; optional MLX Laya; models propose, the app keeps policy)
• Tongyun1/Jev-in-the-Loop §146 (HEAD 039c2117f4e3; operating a browser is not a calibrated Noul)
• hf:aimeigaoshou/agent-jev §146 (sha 024a68eade83; verified:false accuracy 0.7925; Qwen3-0.6B is not Archer)
• hf:clduab11/jev-calibration-statistics §145 (HTTP 200, sha 9bbe055ee875; 0.612 against 0.740 theirs; Gemma 4 is not Archer)

73 novel + 6 revisit. invented_signal: false. Evidence archive: research/archive/hourly/2026-09-21T18/. README still ends at ## License.

Test plan

• python3 .agents/skills/augustus/scripts/uniqueness_gate.py after the §148 renumber (includes fingerprint self-test)
• Re-ADV: occupancy vs glance §147, README License, sample spot-check
• Parent merges only after ADV_PASS

  

[cursor]

ADV review of draft #74 (hourly 1203). Draft stays open. Do not merge this PR in this run.

#73 conflict (section number only)

Open draft #73 (#73, cursor/fold-glance-vlm-logit-140d) also targets main @ 867628040f51 and claims notes.md §147 for yoheinakajima/glance alone. Its fold id is only notes.md §147. It does not claim composition 761-776 or findings batch #129.

#74's §147 does not mention glance, glance-vlm, or yoheinakajima/glance. All 79 census ids are in §147 (73 novel + 6 revisit). Collision is the section number, not a shared card.

No duplicate first-fold of glance. After rebase, #73's glance card stays the one canonical card.

Recommended merge order: merge #73 first (single-peer glance keeps §147). Then rebase #74 onto post-#73 main and renumber this fold's notes section from §147 to §148, including the uniqueness lock token notes.md §147, the gate heading check, and run_digest.json notes_section. Keep composition 761-776 and findings batch #129. #73 does not use those ids.

Honesty gates

• Soft scores are marked theirs and are not hard gates. laya-ft false alarms are not a license to auto-block. jev-gate noul p=0.02 is not a merge. arbitype gate is still a soft score until policy sets the block. muse-skills shadow mode proceeds anyway. No new Augustus runtime call site wires these peers into the skill.
• invented_signal: false in notes §147 and research/archive/hourly/2026-09-21T18/run_digest.json.
• REVISIT densifies stay on the prior section ids. jev-seo stays §131 (HEAD f42455ac951a was f8cb7c55c356). jevnav stays §129 (b7a12d2f54ce was 96f5438bea96). Jev-in-the-Loop and agent-jev stay §146. jev-calibration-statistics stays §145 (HTTP 200 was 401). No sibling first-sighting headings. SoundBlaster/SwiftDecision is a first card: revisit tag, no prior notes card. SwiftDecision-Examples is a separate novel remainder.
• README still ends at ## License. No uniqueness lock or HIGH wall after License.
• CHANGELOG.md has no 1203 uniqueness dump. The lock lives in research/changelog-hourly.md, research/notes.md, and the other research overlays. uniq_1203.txt is under research/archive/hourly/2026-09-21T18/.
• Framing is the decision-model class. TypeSafe Jev (Choice, Score, Noul) is the dominant exemplar and the default path, not the whole mandate.
• Added human-facing lines in README, SKILL, notes, and composition have no new em dashes.
• python3 .agents/skills/augustus/scripts/uniqueness_gate.py passed here: 1203 chars=11524.

Occupancy (against merged main)

• notes §147 is the next free section after merged Fold hourly 1110 HIGH: laya-vision same-species serving + novel cards + REVISIT densify #72 §146. It does not reclaim §146.
• composition 761-776 follows 760. It does not reclaim 745-760.
• findings batch #129 is new above batch #128. It does not reclaim #128.
• Those three ids are free on current main. The only concurrent claim is §147 from Fold yoheinakajima/glance (local VLM logit harness) #73.

Spot-recapture (live upstream)

No vision peer in the 79. lightjev says no screenshots. jevnav densify is page truth, not pixels. The vision peer for this hour is yoheinakajima/glance, and it is only in #73.

Gate-adjacent and other novel HIGH, checked live:

1. peterfriese/jev-foundation-models (PRIMARY). HEAD 27963995965d and README blob b40b836f4396 match. Apache-2.0, Swift. README: 40-150 ms, zero hallucinations, calibrated probabilities, Never embed API keys, @Generable onto noul/choice/score. Those sentences are theirs. Stars are 2 now versus 1 on the card. Star-noise after the pin, not a content miss.
2. Alexander-Ollman/laya-ft (safety / auto-block adjacent). HEAD 41f8247c3969, README e97e314c88d5, 0 stars, license unset. Live table matches: Aegis prompts F1 60.6% / 81.5% / 83.7%, BeaverTails 6.0% / 34.1% / 59.4%, XSTest 250 harmless prompts, false alarms 24.8% to 65.2% to 47.2%, 67,890 decisions, No Jev outputs were used, not a dependable general-purpose safety filter.
3. ruslanlap/jev-gate. Pinned commit 300014a1bdea exists (2026-09-21T18:10:43Z) and the card's README blob 96e31d8dc5d8 is that tree. Live default branch has since moved to 2489de2c1ee0 (pushed 18:19Z). On the live README the quoted figures still hold: measured $0.000143, verdict p=0.02, probabilities must sum to 1 within 0.02, exit code is code. Pin is honest. Post-pin SHA move is not a new Harbor.
4. Renwang-Huang/arbitype. HEAD 454cf2c4a345, README ffa11f281a52. README: independent open-source project, not an official TypeSafe AI product. Classify, score, verify, gate, route, review. PyPI badge is 0.6.0 pending. Gate result is a signal, not proof.
5. hfnissum-byte/jevmerge. Pinned commit 2b472ca7304b (README b32e85e4909c) says code computes candidate resolutions, drops ones that do not parse, the model picks, code gates the pick. 1 star. Live HEAD 25af4c00c663 (18:17Z) retitles the README to Hunkpick while the GitHub repo name is still jevmerge. The enumeration claim is true at the pin.
6. Ejokey/lightjev. HEAD 921fffb588b9, README 30fa74b9f711. Six pages, four hits, target 3/3, $0.000851, 23.6s, STOP always available, about $0.0001 per decision. One crawl is not Harbor.
7. SoundBlaster/SwiftDecision (first card). HEAD 7af9416e1ac4, README 54ad5d8e4886, 0 stars, Apache-2.0. README: Models propose. Your application keeps control of policy and action. Optional MLX Laya is a backend, not a replica.
8. theglitcharchitect/muse-skills. HEAD f0cc9cc9cea6, README a14927d3b12f. Shadow mode: the action proceeds and the decision is logged. Shadow is not a gate.
9. hf:clduab11/jev-calibration-statistics (densify §145). HTTP 200, sha 9bbe055ee875, 0 likes, mit. README: 0.612 against 0.740, missed its main pre-registered bar, AUROC 0.899 on 9,075 passages from 349 questions, jev-1.13.0, Gemma 4 E2B. Gemma 4 is not Archer.
10. hf:aimeigaoshou/agent-jev (densify §146). sha 024a68eade83, 0 likes, apache-2.0. model-index verified false: Top-1 0.7925, Brier 0.0448, ECE 0.1687. Same numbers as §146, not a new Harbor.

Codercise/jev-in-practice and emlama/jev-mcp pins also match (key stays off the client; one SQLite volume).

Verdict

ADV_PASS on the cards. Merge hold: do not merge #74 until #73 is merged and this branch is rebased with notes §147 renumbered to §148. Composition 761-776 and findings batch #129 stay. Glance remains #73's single canonical card. No card rewrite is required for the recaptured quotes.

[24601]

Re-ADV after rebase onto post-#73 main (8a096a8, 1d24465). Draft stays open. Do not merge in this run.

Occupancy vs glance §147

Glance (merged #73) keeps notes.md §147. Hourly 1203 is ## 148. Hourly 1203 HIGH. The 1203 section points at glance and does not re-card it. Composition stays 761-776 (glance added none; main still ended at 760). Findings batch stays #129 (glance added none). Fingerprints: the glance look stays notes_section 147; 70 hourly-1203 looks moved to 148. Revisit densifies stay on §145 / §146 / §131 / §129.

uniqueness_gate

python3 .agents/skills/augustus/scripts/uniqueness_gate.py passed: 1203 chars=11524. Glance lock still ends notes.md §147. The 1203 lock token is notes.md §148.

README License

README still ends at ## License. No uniqueness dump after License.

Sample spot-check

Card text was not rewritten. Still in §148: 40 to 150 ms, Never embed API keys, false alarms 24.8% to 65.2% to 47.2%, 67,890 decisions, No Jev outputs were used, $0.000143, sum to 1 within 0.02, $0.000851, 0.612 against 0.740, verified:false accuracy 0.7925, Models propose / application keeps policy, shadow mode proceeds anyway. Soft scores stay theirs and are not hard gates. invented_signal: false.

Verdict

ADV_PASS. Parent may merge. This run does not merge.