Fold hourly 2347 HIGH - #59
Conversation
Densify razorback16/openjev MLX serving port and TypeLLM Release v0.1.1 on the original cards. Fold novel class members (JevLoop, option-order, jevtok, ockev, n=8 pilot) with *theirs* honesty locks. Does not bump 0.5.0. Co-authored-by: Basit Mustafa <24601@users.noreply.github.com>
|
Testing gate: PASS Independent adversarial review + tests on head Prior twenty-one uniqueness locks (0843–2246 + Open-Jev densify) + REVISIT_LOCK are byte-identical vs
IDs. Claimed Adversarial
Notes, not FAILs. Star-noise after snapshot: openjev 228★ vs fold 225★; simple-jev 415★ vs fold 408★. TypeLLM 26★ still matches. JevLoop 6★ still matches. Hub archerhume/4rcherhume HTTP 401 this pass (fold 401). Description/README gitblobs for the densifies still match. JevLoop live HEAD is now Ready for parent review. PR stays open. Do not merge from this agent. |
Off latest
mainafter merged #58 (427dcf4Fold hourly 2246 HIGH). Do not reopen or amend PR #23–#58. Do not bump 0.5.0.Hourly 2347 HIGH fold. Jev is the exemplar of a backend-agnostic class of categorization/scoring/typed-decision models, not the whole mandate. REVISIT densifies original cards:
razorback16/openjevon §75,TypeLLM/TypeLLMon §113. Do not mint sibling first sightings. SHA move is not a replica. Third-party benches stay theirs.REVISIT densify (theirs)
razorback16/openjevMLX Apple-silicon serving port. Live 225★ (star-noise is not the fold). HEADfebf02e88989→2050fdb8280d(MLX backend: steps > 1, think, and text generation). README SHA242a737dba01→d5322e16e565. Prompt SHAbbd013e74630is image Qs. Quote theirs: MLX backend steps>1/think/text gen + image Qs; both backends also serve text generation on an OpenAI-compatible/v1/chat/completions. dual serving is not generate. Hosted Codiv ≠ TypeSafe. wire-compat ≠ logit-equiv. The 1441 "chat 501 on MLX" line is stale this hour; generating text is still generate, not a calibrated Noul.TypeLLM/TypeLLMGitHub Release v0.1.1. Live 26★ (was 23★; star-noise is not the fold). HEAD unchanged8a8b4aefd443. README SHA unchanged9f6dea3a4c8c. Quote theirs: GitHub Release v0.1.1 published 2026-09-21T04:49:02Z; Drop fixed banner height so it scales on PyPI. Constrained AR ≠ calibrated Noul. PyPI packaging ≠ calibrated Noul. type safety does not guarantee factual accuracy. A release tag is distribution, not calibration.Novel HIGH (first-sighting)
zjunlp/JevLoop6★. Independent not affiliated. Zero-deps offline loop. Not a TypeSafe product.NicolaiLassen/open-bonsai-jev≠NicolaiMTLassen/open-bonzi-jev. WANLI-256 74.6% theirs.novaleolin/jev-evolve(0.188 or 0.542 theirs; option order can change an answer);LabGuy94/jevtok0 mismatches theirs not Harbor;fancyboi999/ockev35ms 95.8% TomatoEggBench theirs not Harbor;zhengbangbo/structured-decision-benchn=8 theirs; n=8 is not Harbor.danielhirt/jev-lab≠ tanayvasishtha/jev-lab ≠ dairui1/jev-lab ≠ mjyoke1111/jev-lab.JairajSustained/llm-routing-jivdoes not execute;s-hiraoku/jev-page-checkeradvisory does not block; compaction plugins ranking before lossless condensation;1deat0r/JcuaCua-S1 ≠ TypeSafe;2951461586/Jev-Register-Toolcatalog only (no key-farming how-tos).sriramkasyap/laya-apiwire-compat ≠ logit-equiv;Yang-SS-stack/jev-computer-use≠ Mrchen116/jev-computer-use;amoreX/jevvy≠ PanAchy/jevvy ≠ aboisvert/jevvy;smile-magic/laya-mlx-ddz≠ smile-magic/laya-mlx-wzq;hf:space:Yuki131/KaLM-Jev≠ KaLM-Embedding/KaLM-Jev (reranker ≠ Jev replica);hf:soyelmismo/laya-multilingual-onnxserving substrate ≠ calibrated replica.nexibeo/jev-cookbook(§70),leesk212/JEV-CPU(§64),kazuhideoki/jev-search(§55). Do not mint siblings.s5juxbumjm/layacmempty SHA;vlasvar/jev-researchplaceholder MVP.Skip Archer (
promised_not_landed). Hub archerhume/4rcherhume HTTP 401.IDs
notes.md§135 (densify cards stay on §75 / §113)Class recipes (not Jev-only)
Without Augustus: treat MLX text gen as a calibrated replica, a GitHub Release as a Noul, n=8 as Harbor, option-order 0.188 as gold, or Cua-S1 as TypeSafe Jev.
With Augustus: dual serving is not generate; Hosted Codiv ≠ TypeSafe; wire-compat ≠ logit-equiv; Constrained AR ≠ calibrated Noul; PyPI packaging ≠ calibrated Noul; n=8 is not Harbor; option order can change an answer; ranking before lossless condensation is still ranking; catalog ≠ endorsement; theirs not Harbor. Same split for any Choice/Score/Noul-style head.
Honesty: theirs on bench numbers; catalog ≠ endorsement; soft judgment never sole veto.
Does not bump 0.5.0. Skip Archer.
invented_signal: false.