Fold hourly 1110 HIGH: laya-vision same-species serving + novel cards + REVISIT densify

[24601]

Summary

Hourly 1110 fold off main at 6be01e0429ed (merged #70). Notes §146, composition 745-760, findings batch #128. Does not bump 0.5.0. Does not reopen #23-#70.

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.

PRIMARY: hf:thaitea/laya-vision (13 likes, sha a2653db2831b, CC-BY-NC-SA-4.0). SmolVLM-256M typed vision decisions (choice / score / noul, one forward pass, no text generation). Same-species serving, not an 18th scoring row. The act/escalate head is untrained (zero gradient, random init). Do not gate on it. A-OKVQA 63.1% ECE 0.266 to 0.094, ScienceQA 89.0%, VQAv2 noul 73.2%, all n=8,235 75.9% calibrated ECE 0.035 are theirs, not Harbor. The VQAv2 split is a re-split, not published VQAv2. Usage still loads thaitea/laya-vision-smolvlm-256m (already §87).

Also carded: AntonG87/codearia-sieve (page to typed fields; parse is not a decision; 6 of 6 and 11 of 11 theirs), 7Zenox/gemma-jev (letter-slot logits; JSON chat is not a calibrated Noul; Gemma and Qwen3.5 are not Archer), Pdbz199/local-decision-model (one pass from the public post; schema-valid is not the same as correct), musubi-labs/musubi-jev (README is a kev tree copy; copied kev numbers are not a musubi bench), quaeast/vllm2jev, IAmJSD/pg-laya, gbesse/jev-rerank-server (ranking is not calibration), MarkChu-git/typesafe-mcp (act_above 0.8 is still soft; soft judgment is not a sole veto).

REVISIT densify, marked theirs, prior section kept:

• pythongiant/laya-drift §142 (description rewrite to monitor agent drift)
• turenlabs/jast §145 (SHA unchanged; star 0 to 1 is star-noise)
• Adkid-Zephyr/chinese-workflow-decision-bench §145 (64/64 and 63/64 theirs; synthetic, not a group-chat dump)
• JabbaKadabra/SystemOneDotNet §143 (unofficial .NET client; wire-compat is not logit-equiv)
• arnavm-codes/JevFence §145 (block table is policy; soft scores are not hard gates)
• cloudbtl/JevRAG first card (revisit tag had no prior notes card; combinatorics is not a bench)
• gbesse/decision-workbench §139 (human review stays separate from model output; demo scores are not accuracy measurements)

69 novel + 7 revisit. invented_signal: false. Evidence archive: research/archive/hourly/2026-09-21T17/.

Test plan

• python3 .agents/skills/augustus/scripts/uniqueness_gate.py (includes fingerprint self-test)
• Adversarial review of quoted theirs numbers against the Hub card and READMEs
• Confirm README still ends at ## License and holds no uniqueness dump

  

[24601]

Adversarial review (hourly 1110, draft PR #72)

Do not merge from this comment. Catalog densify holds. No blockers.

Gates

• Occupancy. Current origin/main is 6216705 (merged #71, Release v0.5.1) and still ends at notes.md §145 / composition items 729-744 / findings batch #127. This fold is §146 / 745-760 / batch #128. No collision with already-merged ranges.
• Release pins. Diff vs base 6be01e04 does not touch README, CHANGELOG, CITATION, marketplace, or docs/release-notes-v0.5.1.md. A three-way merge onto 6216705 stays clean and keeps version: 0.5.1, the release notes, and the slim skill, then adds the 1110 lock. Fold text "Does not bump 0.5.0" is this fold's non-bump, not a pin rewrite.
• README / uniqueness. README.md still ends at ## License (MIT / SECURITY / CONTRIBUTING only). python3 .agents/skills/augustus/scripts/uniqueness_gate.py prints uniqueness-gate ok. invented_signal: false in research/archive/hourly/2026-09-21T17/run_digest.json. Added human prose has no U+2014.
• Class framing. TypeSafe Jev stays the default recommended path. The lane stays the wider decision-model class. laya-vision is same-species vision serving, not an 18th scoring-table row (no scoring-table edit). Act/escalate head is untrained (zero gradient, random init) and is not a gate.
• musubi-jev. README at e943f21e4057 is the kev tree (# Kev, jaredpalmer CI badge). The card says copied kev numbers are not a musubi bench and does not paste a kev table as a musubi measurement. musubi-labs/musubi-jev ≠ jaredpalmer/kev.
• Soft scores. act_above 0.8 is still soft. JevFence: soft scores are not hard gates. Soft judgment is not a sole veto. Counts: 69 novel + 7 revisit.

theirs checked live

• thaitea/laya-vision sha a2653db2831b, CC-BY-NC-SA-4.0, 13 likes. A-OKVQA 63.1% ECE 0.266 to 0.094 (n=1,138). ScienceQA image 89.0% ECE 0.080 to 0.034 (n=2,097). VQAv2 noul 73.2% ECE 0.085 to 0.042 (n=5,000, re-split, not published VQAv2). All n=8,235 75.9%, calibrated ECE 0.035 (raw ECE on that row is 0.108; the fold's 0.035 is the calibrated figure). About 71 ms. Option order varies 1.2 points. Train 94.7% vs val 63.1% on A-OKVQA. Usage loads thaitea/laya-vision-smolvlm-256m (§87 card id, not this card).
• 7Zenox/gemma-jev 144 authored decisions. Gemma 3 270M 0.293 below chance 0.333. Gemma 4 E2B-it JSON chat 0.807. Qwen3.5-4B JSON chat 0.813. bf16 vs fp32 argmax-agreement check not run. JSON chat and letter-slot softmax are not a calibrated Noul.
• AntonG87/codearia-sieve 6 of 6 and 11 of 11 are the judge table, marked not Harbor. Parse is not a decision.
• musubi-labs/musubi-jev kev copy, not a new bench.
• Pdbz199/local-decision-model independent from the public post. The 70 to 500 ms row is attributed to the Jev post, not measured here. Span start/end head is locate, not decide.
• gbesse/jev-rerank-server SciFact n=25 nDCG@10 0.616377 to 0.718260, Recall@10 0.84 unchanged. Ranking is not calibration.
• MarkChu-git/typesafe-mcp act/review/abstain is code. Default act_above 0.8, review_above 0.5, still soft.
• Adkid-Zephyr/chinese-workflow-decision-bench 64/64 and 63/64, synthetic, not a group-chat dump. 8 cases times 2 prompts, CPU and MPS labels matched, and that match is not a pretraining proof.
• pythongiant/laya-drift description is now monitor agent drift (was calculate agentic drift, §142). alignment 0.7 and plan_ref 0.3 are application policy. Densify, not a sibling.
• cloudbtl/JevRAG 20 branches times six hops is combinatorics, not a bench. walk, gather, file, watch. First card.
• quaeast/vllm2jev does not reproduce Jev calibration. wire-compat is not logit-equiv.
• malevrigns/agent-jev 79.25% (1585/2000), ECE 0.1687, marked theirs not Harbor.
• DjTaNg-404/Laya-2048 one seed-48 game 28,092, marked not Harbor. jast HEAD c6588285208f and README SHA d9214ca31f92 match §145; star 0 to 1 is star-noise.

Non-blocking

The Hub limitations also say score questions are untrained (no ordinal image data, outputs meaningless). The fold quotes the card lead (choice / score / noul) and the measured choice and noul rows, and it refuses the act head. It does not quote a score accuracy. Composition item 754 puts 64/64 on the line after the laya-drift sentence; the item title and the uniqueness lock attribute those fractions to the chinese bench, and §146 does too.

ADV_PASS