Fold hourly 1843 HIGH - #52
Conversation
Archive digest still claimed §128 / 465-480 / #110 after the rebase onto merged 1746. Point it at the free IDs §129 / 481-496 / #111 and gate those fields in uniqueness_gate so a stale digest fails the lock. Co-authored-by: Basit Mustafa <24601@users.noreply.github.com>
|
Testing gate: PASS Independent adversarial review + tests on head Pushable FAIL fixed on branch (
IDs. Claimed Adversarial (PASS after digest remap)
Notes, not FAILs. Star-noise after snapshot: kev 1051★ vs fold 1033 (SHA still Ready for parent review. PR stays open. Do not merge from this agent. |
Hourly 1843 HIGH fold off latest
mainafter merged #51 (1746 / §128). Do not reopen or amend PR #23–#51. Do not amend #51.What this hour is
Primary is a REVISIT densify of
jaredpalmer/kevon the original §45 card: own-data JSONL fine-tune and--init_fromwarm-start LoRA/head (PR #9). Jev is the exemplar of a backend-agnostic class of categorization/scoring models, not the whole mandate. SHA move is not a replica.dabit3/jev-experimentsandsimota/tenbinstay densify-not-sibling.kev own-data JSONL (densify §45)
HEAD
bd058057ad0a(was8465c4c4c294/ 1746). README SHA84b872488915. Quote theirs: Fine-tuning on your own data.--dataJSONL.--init_fromwarm-start LoRA/head PR #9. Kev-0.8B / 4B / 9B Qwen3.5 family. 4B new-source 0.794/0.832 theirs. 9B new-source 0.812/0.837 theirs. From-scratch on 836 support-tool decisions scored 0.33 against 0.84 for the released model;--init_fromkept 0.83 there and reached 0.88 on the new domain theirs. Kev-0.5B card now points at the Qwen3.5 family. No Jev outputs were used for training. option order can change an answer. 8.2% ≥0.9 on wrong theirs (Kev-9B 7.5%).from-scratch ≠ warm-start. JSONL labels ≠ Harbor. Qwen3.5 ≠ Archer. wire-compat ≠ logit-equiv.
Novel HIGH (first-sighting)
kyegomez/open-jev: unofficial research implementation with random weights. reconstruction ≠ replica. ≠ razorback16/openjev ≠ TheoLeeCJ/openjev ≠ Zefan-Cai/Open-Jev ≠ Shalimov04/open-jev.jourdanlabs/assay-001: split verdict. CLINC150 ECE 0.0204 theirs. Banking77 ECE 0.0936 theirs. 8,576 responses zero type errors theirs. Not a global certificate.brnyxx/jev-ra: 3-5x / ~300 ms theirs. 8.50× Wikipedia theirs. systems comparison ≠ semantic equivalence.Promethe-us/awesome-jev: catalog ≠ endorsement. ≠ MrJev/awesome-jev ≠ yibie/awesome-jev.ThePFMind/jev-mcp: ≠ jkudish/jev-mcp ≠ burnigtm/jev-mcp. serving substrate ≠ calibrated replica.Remainder: minesweeper cousins (game success ≠ calibrated Noul),
namenu/pi-jev-effort≠TheoOliveira/pi-jev,samatv256/mini-Jev≠r-ms/mini-jev, HF allmix vs handoff-demo, Nutlope/jev-fraud, jeffloo886/jev-notion. Skip Archer (promised_not_landed). Hub archerhume/4rcherhume HTTP 401.IDs
notes.md§129Class recipes (not Jev-only)
Without Augustus: treat a from-scratch adapter as equal to a warm start, a JSONL file as Harbor, a random-weight reconstruction as a replica, or a catalog as a grant.
With Augustus: from-scratch ≠ warm-start; JSONL labels ≠ Harbor; reconstruction ≠ replica; catalog ≠ endorsement. Same split for any Choice/Score/Noul-style head.
Does not bump 0.5.0. Skip Archer.
invented_signal: false.Checks (this revision)
Head
60b6746remaps the 1843 archiverun_digest.jsonoff #51's IDs (§128 / 465–480 / #110) onto the free IDs (§129 / 481–496 / #111). uniqueness_gate now refuses a stale digest.uniqueness_gate.pyok (1843 lock 1815 chars, one substring in 21 overlays; 1746 lock retained; CHANGELOG has no dump wall)evaluate_decisions.py --self-testok (from-scratch ≠ warm-start / JSONL labels ≠ Harbor / reconstruction ≠ replica, plus 1746 helpers)revisit_fingerprints.py --self-testokyaml.safe_loadok