0823 fold: novel HIGH + REVISIT densify

[24601]

Hourly 0823 HIGH fold off latest main after merged #66 (0707 / §142). Fresh PR. Do not reopen or amend PR #23–#66. Does not bump 0.5.0.

What this folds

Novel HIGH first cards (live REST this hour):

• FluidInference/FluidUse 3★ Apache-2.0 Swift: on-device Mac form CU via Accessibility + CUA-S1-FORMS CoreML (~706K params, ~1ms Neural Engine). Quote theirs: field→value match among supplied options not free text. Cua-S1 ≠ TypeSafe. ≠ FluidInference/laya-coreml ≠ AndyInQtr/laya-coreai ≠ mizorewww/laya-coreml. serving substrate ≠ calibrated replica.
• shhivv/third-hand 274★ MIT Swift: Quote theirs: Screenshots aren't uploaded. Jev is the only model. not fully offline.
• typesafe-ai/system-one-adapter-python 226★ MIT: Drop-in TypeSafeClient replacement backed by LLM APIs. wire-compat ≠ logit-equiv.
• typesafe-ai/typesafe-sdk-js 203★ MIT and typesafe-ai/typesafe-sdk-python 175★ MIT: first-card GitHub faces (already one-liners in §4). catalog ≠ endorsement.
• r-ms/mini-jev 40★ MIT: Quote theirs: read the option letter's logits instead of generating JSON. JSON 0.909 letters 0.907 theirs. 13 600 / 13 600 theirs. softmax over letters ≠ calibrated Noul. ≠ yuki-oshio/mini-jev.
• Das-rebel/a3m-router 16★ MIT: model=jev-auto. routing ≠ permission.
• AbdelStark/jev-benchmarks 13★ Apache-2.0: AG News 0.910 theirs. Banking77 0.870 theirs. DAIR Emotion 0.480 theirs. theirs not Harbor.
• Remainder first cards: siliconkernel/vllm-jev-decison (No generative fallback), Stumble/jev-go, Twister915/typesafe-ai, rorshopping/jev-on-a-laptop (Unofficial research repo. Not affiliated with TypeSafe AI), TypeSafeAI/typesafe-playground, nickthompson480/typesafe-ai-playground, plus named remainder and two HF corpora.

REVISIT densify (keep prior section ids; no sibling first sightings):

• PsiACE/dohnuts PRIMARY densify §137: HEAD 253766e5fcb7 README 4b5b019deba9 (was a5049834489c / e1b448c440b5). 11★ Apache-2.0. MODEL_CARD + training dataset refs. JevBench 65.80% vs Jev 86.58% / Laya multi 47.62% theirs. 152 / 231 theirs. 78.21% macro accuracy theirs. 180,031 decisions theirs. Qwen3.5-0.8B. Joint RLCD theirs. Dohnuts ≠ TypeSafe. same-species serving not an 18th scoring row. densify §137 not a sibling first sighting. SHA move is not a replica.
• browser-use/jev-ultrafast densify §4 description rewrite: HEAD 1231850a0bf1 README fa7d079f9192. Quote theirs: Fastest and cheapest web agent.
• pythongiant/laya-drift densify §142: HEAD 334953e5cb8f README 6662121ba0f3 (was fc94b71cf7dd / 55ef2343ee5d). Quote theirs: opencode plugin to calculate agentic drift over time. densify §142 not a sibling first sighting.
• BlinkWrite/pii-masker first card (revisit tag, no prior notes card): On-device reversible PII masking theirs. Locate ≠ decide.

Skip-thin empty SHA: GokhanCalkap/LayaCode, fredzhaozonghui/LAYA1. Futureppo/typesafe_register key-farming skip. Skip Archer.

IDs

notes.md §143 / composition 697–712 / findings #125 / archive research/archive/hourly/2026-09-21T14/.

Checks

• uniqueness_gate.py ok (0823 lock in 21 overlays; CHANGELOG has no dump wall)
• evaluate_decisions.py --self-test ok
• revisit_fingerprints.py --self-test ok
• invented_signal: false

Catalog ≠ endorsement. Third-party benches stay theirs.

  

[24601]

ADV_PASS. Adversarial review + testing of #67 (HEAD 48aeadff1968f47409019f9088879c56b4c7fd8d).

Honesty checked vs live READMEs / MODEL_CARD / GitHub descriptions at claimed SHAs: PsiACE/dohnuts densify PRIMARY 11★ Apache-2.0 HEAD 253766e5fcb7 README SHA 4b5b019deba9 was a5049834489c / e1b448c440b5; MODEL_CARD + training dataset refs; JevBench 152 / 231 65.80% vs Jev 86.58% / Laya multi 47.62% theirs; 78.21% macro accuracy theirs; 180,031 decisions theirs; Qwen3.5-0.8B; Joint RLCD theirs; Dohnuts ≠ TypeSafe; theirs not Harbor (JevBench 65.80% is not a hard gate). FluidInference/FluidUse first card 3★ Apache-2.0 Swift HEAD c18071d791eb README SHA f1cba4a244bb; on-device Mac form CU; Accessibility API; CUA-S1-FORMS CoreML ~706K params ~1ms Neural Engine; quote theirs The model matches a field to a value it is given. It does not write free text; field→value among supplied options only; Cua-S1 ≠ TypeSafe; FluidInference/FluidUse ≠ FluidInference/laya-coreml ≠ AndyInQtr/laya-coreai ≠ mizorewww/laya-coreml; serving substrate ≠ calibrated replica.

Also first-carded / densified (quotes live): shhivv/third-hand 274★ MIT HEAD 430394b35dbb README SHA b615d7c3fd19 Screenshots aren't uploaded; Jev is the only model; not fully offline. typesafe-ai/system-one-adapter-python 226★ MIT HEAD adffc2eab300 README SHA d01afbf0499e Drop-in TypeSafeClient replacement backed by LLM APIs; wire-compat ≠ logit-equiv. typesafe-ai/typesafe-sdk-js / typesafe-sdk-python GitHub faces; catalog ≠ endorsement. r-ms/mini-jev 40★ MIT HEAD ca612198bfb6 README SHA 565b70c4cf4d JSON 0.909 letters 0.907 theirs; 13 600 / 13 600 theirs; softmax over letters ≠ calibrated Noul; yuki-oshio/mini-jev ≠ r-ms/mini-jev. Das-rebel/a3m-router 16★ MIT HEAD 62caefe59315 README SHA c19e802d5cf2 model=jev-auto; routing ≠ permission. AbdelStark/jev-benchmarks 13★ Apache-2.0 HEAD 0d610cc53e79 README SHA 5fd3627f7de4 AG News 0.910 / Banking77 0.870 / DAIR Emotion 0.480 theirs not Harbor. siliconkernel/vllm-jev-decison No generative fallback. rorshopping/jev-on-a-laptop Unofficial research repo. Not affiliated with TypeSafe AI. Remainder first cards named, not a dump wall.

REVISIT densify kept prior section ids (no sibling first sightings; no reopened uniqueness walls): PsiACE/dohnuts densify §137 in-place (fingerprint notes_section 137); same-species serving not an 18th scoring row (0049 JevBench public-subset table not extended). browser-use/jev-ultrafast densify §4 description rewrite HEAD 1231850a0bf1 README SHA fa7d079f9192 GitHub description Fastest and cheapest web agent theirs; fingerprint notes_section 4. pythongiant/laya-drift densify §142 HEAD 334953e5cb8f README SHA 6662121ba0f3 was fc94b71cf7dd / 55ef2343ee5d GitHub description opencode plugin to calculate agentic drift over time theirs; fingerprint notes_section 142. BlinkWrite/pii-masker first card (revisit tag, no prior notes card) HEAD 6ad202ab4443 README SHA 27931758af6c On-device reversible PII masking theirs; Locate ≠ decide; fingerprint notes_section 143. First-party 0707 SHA bumps are not densify REVISIT (kev fingerprint still 4f8110a3f862 last_look 11:51Z / notes_section 45). UNIQ_0707 / UNIQ_0151 / UNIQ_0551 / UNIQ_0445 strings identical to main. Soft judgment ≠ hard gate throughout. Archer still promised_not_landed (Hub 401). invented_signal false. No version bump. Do not reopen #23–#66.

Tests (python3, all exit 0):

• .agents/skills/augustus/scripts/uniqueness_gate.py → uniqueness-gate ok (0823 chars=3579, 21 overlays, 5 revisit overlays)
• .agents/skills/augustus/scripts/evaluate_decisions.py --self-test → self-test ok
• research/revisit_fingerprints.py --self-test → revisit-fingerprints self-test ok

IDs unique vs main (41fa40c / merged #66): §143 / items 697–712 / batch #125 (main owns §142 / 681–696 / #124). Occupancy steal FAIL-as-expected for §143, 697, 712, #125; this branch only. SKILL 0.5.0 frozen. CHANGELOG Unreleased has no uniqueness dump wall. CI build success. READY_TO_MERGE. Do not merge from this comment.