Densify Open-Jev: latency, providers, TREC, HF packs

[24601]

Revisit densify of Zefan-Cai/Open-Jev on the original hourly 1441 card (notes.md §125). Rebased onto latest main after merged #54 (6a7a557, aisearchio census gaps / §130 / items 497–504 / batch #112). Do not mint a sibling first sighting and do not claim §130. Do not reopen or amend PR #23–#52. Does not bump 0.5.0.

What this look is

Material since last look is README / latency report / provider comparison / TREC-DL / public HF packs / official launch X thread, not the 0★→3★ star move. SHA move is not a replica. Independent of TypeSafe. No RLCD/parity claims.

HEAD 4933ee84951f (was 6d8de5ed72a0). README SHA ce1a587219e4. Pushed 2026-09-21T01:34:47Z. Astra TREC commit 1dd56990be7e (pushed 2026-09-21T01:17Z). Live 3★ at fold (star-noise is not the fold).

Honesty locks

• LoRA adapters plus a trained scalar decision head and calibration temperature. not merged base models. LoRA ≠ RLCD replica.
• Public HF packs (revisions unchanged this look): dataset ZefanCai/Open-Jev rev c67699e13d0a; Open-Jev-2B rev 0c7aa498b162; Open-Jev-9B rev 47e966881e48. 27B still in progress.
• Latency vs Jev-1.13.0 theirs: customer-service P50 local HTTP 85.03 ms versus Jev HTTPS 295.26 ms. At 1024 tokens / 32 candidates Open-Jev is slower: 1015.90 ms versus 301.37 ms. Hardware and network paths differ. systems latency ≠ semantic equivalence. Prefix caching experimental/off by default.
• Provider comparison theirs: GPT Luna P50 918.13 ms, Astra 1938.39 ms. Latency does not establish equal task quality.
• TREC-DL theirs: Jev / Luna / Astra completed. Open-Jev TREC pending. Do not quote Jev/Luna/Astra nDCG as an Open-Jev result.
• Training restated theirs not Harbor: 80,816 rows; 2B 94.71% / OOD 86.02%; 9B 97.54% / 91.97%. hard acc ≠ calibrated Noul. type-valid ≠ exact. Qwen/Qwen3.8-27B ≠ Archer.
• Website https://zefan-cai.github.io/open-jev/ plus launch X thread theirs. Quote theirs: Inspired by Jev, we built Open-Jev: open-source decision models. 2B/9B: LoRA adapters + decision heads. Demo. Intro reply (52-second intro; HF collection ZefanCai/open-jev-6ab049b9d43a267bae4dedc8). Website reply. Zefan-Cai/Open-Jev ≠ TheoLeeCJ/openjev ≠ razorback16/openjev ≠ Shalimov04/open-jev ≠ kyegomez/open-jev.

IDs

• Densify original notes.md §125 (items 425–426 stay in place). Merged Fold aisearchio census gaps HIGH #54 owns §130 / items 497–504 / batch #112. Next free IDs §131 / 505+ / #113 are reserved.
• uniqueness_gate now checks 0843 + 0915 + jcr + 0922 + 0940 + 0947 + 1049 + 1143 + 1248 + 1340 + 1441 + 1542 + 1643 + 1746 + 1843 + 1936 + Open-Jev densify (21 overlays; both UNIQ_1936 and UNIQ_OPENJEV as consecutive substrings)
• Evaluator: systems latency ≠ semantic equivalence / hard acc ≠ Noul / LoRA pack ≠ merged base / prefix cache off / TREC pending / Qwen/Qwen3.8-27B ≠ Archer / latency ≠ quality
• Fingerprints: github:Zefan-Cai/Open-Jev SHA 4933ee84951f / README ce1a587219e4; HF 2B/9B packs + dataset last_look; densify_original_ids lock these ids to §125

Quote theirs. Skip Archer. invented_signal: false. Do not merge this PR from this comment.

  

[cursor]

Rebased onto latest main (6a7a557, merged #54 / §130). Head 72524817682d6fed2112dbe4b5c120b7d86733d0. GitHub reports MERGEABLE / CLEAN.

Densify stays on original notes.md §125. Does not claim §130 or mint a sibling first-sighting. Honesty locks unchanged (LoRA + scalar head, theirs latency, HF pack revs). Launch tweet details (website, HF collection, demo, quote) are on the §125 card. Does not bump 0.5.0. Not merging.

[24601]

Testing gate: PASS

Independent adversarial review + tests on head 72524817682d6fed2112dbe4b5c120b7d86733d0 vs origin/main 6a7a557 (merged #54). GitHub: MERGEABLE / CLEAN. Pages build SUCCESS on this head. No pushable FAILs. Do not merge from this agent.

Nothing fixed on branch. Prior sixteen uniqueness locks (0843–1936) + REVISIT_LOCK are byte-identical vs origin/main. CHANGELOG [0.5.0] identical vs main. Occupancy loop range(497, 505) identical vs main (497–504 inclusive). Deleting item 497 fails locally (proved; restored).

#	Check	Result	Evidence
1	Checkout head; SHA; mergeability	PASS	HEAD 72524817682d6fed2112dbe4b5c120b7d86733d0. gh pr view 53 --json mergeable,mergeStateStatus → MERGEABLE / CLEAN. Base 6a7a557496bad773bdee087f81a12245ec4b5c61. merge-base is origin/main.
2	Diff vs main includes merged #54; densify stays on §125; does not reclaim §130 or collide with #54 IDs	PASS	origin/main occupies §114–§130 / composition through 504 / findings through #112. This fold adds no new ## heading in notes.md (no §131). Composition added items []; max still 504; unused 317–321 still absent; no 505+. Findings added batches none; densify subsection is under batch #107 / §125. Items 425–426 densified in place. Overlay lock dumps sit after the #112 lock (same overlay pattern), they do not mint #113 or rewrite #54 cards.
3	uniqueness_gate / evaluate_decisions / revisit_fingerprints; occupancy / hashes / SHAs	PASS	All three commands exit 0 (output below). uniqueness_gate: 21 overlays; UNIQ_OPENJEV new (1545 chars); UNIQ_1936 still required. Fingerprint store: github:Zefan-Cai/Open-Jev full HEAD 4933ee84951f1a3b94b8be1f7490f02a4fa4ba24 / README ce1a587219e4b9bcb5db3787a654f248d9a84b92; HF 2B 0c7aa498b162… / 9B 47e966881e48… newly seeded (absent on main, notes_section 125); dataset SHA unchanged c67699e13d0a…. densify_original_ids locks GitHub + HF 2B/9B + dataset to 125. No duplicate look ids. Live REST this pass matches those SHAs.
4	SKILL metadata stays 0.5.0; marketplace 0.5.0; no release bump	PASS	YAML metadata.version: 0.5.0; typesafe_skill: v0.5.7 @ 65a39f3 (identical vs main). uniqueness_gate refuses any other pin. .claude-plugin/marketplace.json 0.5.0 identical vs main (not in the diff). CHANGELOG [0.5.0] intact; Unreleased is a short Open-Jev bullet + class recipe, not a bump. No uniqueness lock: dump in CHANGELOG.
5	Honesty: LoRA+scalar ≠ merged base; theirs latency; independent of TypeSafe/RLCD/Archer; TREC pending	PASS	Live README theirs: LoRA adapters plus trained scalar decision head and calibration temperature; not merged base models; independent implementation inspired by TypeSafe's Jev; does not reproduce proprietary RLCD; does not claim TypeSafe speedups or parity. Latency theirs: customer-service P50 85.03 ms vs Jev HTTPS 295.26 ms; 1024/32 1015.90 vs 301.37; Luna 918.13 / Astra 1938.39. Prefix caching experimental and off by default; CUDA prefix caching exceeded tolerance on 9/11. Open-Jev TREC results remain pending (Jev/Luna/Astra completed 97 queries). hard acc 94.71% / 97.54% restated theirs not Harbor. Evaluator: systems_latency_is_not_semantic_equivalence, hard_acc_is_not_calibrated_noul, lora_pack_is_not_merged_base, prefix_cache_off_by_default, open_jev_trec_is_pending, qwen38_27b_is_not_archer, latency_is_not_task_quality, lora_is_not_rlcd_replica. Namesake lock: Zefan-Cai/Open-Jev ≠ TheoLeeCJ/openjev ≠ razorback16/openjev ≠ Shalimov04/open-jev ≠ kyegomez/open-jev. Fold skips Archer rewrite.
6	Human-facing prose without em-dash walls; invented_signal false; Pages	PASS	New CHANGELOG Unreleased Open-Jev block: 0 em dashes (en-dashes only on ID ranges / PR spans). Ecosystem / recipes use the existing gloss pattern, not a dump wall. No invented_signal: true in the diff. refresh-log / notes / findings quote invented_signal: false. Pages build SUCCESS on 7252481 (jekyll-build-pages + gate greps).

IDs. Densify original notes.md §125 (items 425–426 stay in place). Merged #54 keeps §130 / composition 497–504 / findings batch #112. Next free IDs §131 / 505+ / #113 unused. uniqueness_gate still requires §130 heading, items 497–504, and Batch #112.

Adversarial (PASS, no fix)

• uniqueness_gate composition loop range(497, 505) (497–504 inclusive) matches merged Fold aisearchio census gaps HIGH #54. Steal-497 now fails (proved locally; file restored).
• Densify vs new: Open-Jev densifies §125. No sibling ## heading. Store notes_section stays 125. Helper densify_original_ids refuses moving it off 125.
• Live SHA this pass still 4933ee84951f / README git-blob ce1a587219e4 / pushed 2026-09-21T01:34:47Z. Astra TREC commit 1dd56990be7e (pushed 2026-09-21T01:17:15Z). HF pack revs unchanged vs the recorded pins (and vs dataset SHA on main).
• Do not reopen Fold hourly 0042 HIGH: jev-readout, jevify, laya-jev-lab #23–Fold hourly 1843 HIGH #52. No wrappers / keys / npm / pip / uv / docker / hf download recipes in the fold prose (anti-copy only).

Notes, not FAILs. Star-noise after snapshot: live 6★ vs fold 3★ (SHA still 4933ee84951f). CHANGELOG Unreleased Open-Jev bullet lists uniqueness_gate through 1843 + Open-Jev densify and omits 1936; the 1936 Unreleased bullet remains immediately below and uniqueness_gate checks both UNIQ_1936 and UNIQ_OPENJEV. Missing blank line between the Open-Jev recipe and the 1936 heading. Do-not-reopen list in the Open-Jev lock still ends at #52 (lock written before #54 merged; this densify does not amend #54).

uniqueness-gate ok
0843 chars=2423 0915 chars=1132 jcr chars=1499 lock0922 chars=1292 0940 chars=1350 0947 chars=2350 1049 chars=2162 1143 chars=2188 1248 chars=2226 1340 chars=1499 1441 chars=1635 1542 chars=2015 1643 chars=2193 1746 chars=1902 1843 chars=1815 1936 chars=1309 openjev_densify chars=1545 revisit chars=447 overlays=21 revisit_overlays=5
self-test ok
revisit-fingerprints self-test ok

Ready for parent review. PR stays open. Do not merge from this agent.