{ "schema": "boar-publication/experiments", "schema_version": "1.0.0", "generated": "2026-10-25", "license": "CC BY 4.0 (text and data); see /CITATION.cff", "count": 387, "verdicts": { "adopted": "shipped or recommended by its frozen rule", "closed": "rule failed, CI excluded the hoped-for gain, a step-0 gate failed, or the mechanism was falsified", "not-detected": "CI straddles zero below the test's minimum detectable effect: not evidence of no effect", "parity": "an equivalence (same quality, cheaper or transferable), not a gain", "measurement": "measures a property; no lever verdict", "tooling": "method or infrastructure arm", "open": "parked, running or awaiting a result" }, "notes": [ "result_verified = false: the result text comes from the 2026-10-13 ledger audit (arm-index v2) or the registry row and was not re-verified for this publication; render it as \"as recorded\".", "hypothesis is null where the hypothesis text lives only in the criterion file (not extracted in this pass)." ], "experiments": [ { "id": "X-001", "code": "ENWIKI100K-v1", "name": "Pack: enwiki top-100k for the 4B", "family": "4B", "question": "Pack: enwiki top-100k for the 4B", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": null }, "set": "iPhone and Android profiles, exact-key-in-context gate", "judge": null, "n": null, "seeds": null, "result": "+3 (bar +6); retrieval 3.3× slower", "result_verified": true, "effect": "+3 (bar +6); retrieval 3.3× slower", "ci": null, "mde": "n/a (gate)", "verdict": "closed", "verdict_detail": "closed (gate FAIL)", "reopen_if": "#6b named-article cap (running at the time; see QUEUE l.155-156)", "closed_row": "R5", "registry_public": "boar-research:reports/CLOSED.md:20", "date": "2026-10-01", "fact_ids": [ "F-1114" ], "chapter": "walls", "source": "closed registry row R5 + verified fact" }, { "id": "X-002", "code": "AR", "name": "AR", "family": "4B-fast-path", "question": "Does a deterministic arithmetic calculator (rate x quantity -> time, ceil-division) plus a detector numeric guard answer the v0.1 arithmetic items exactly with 0 false fires, vs cut 2 (no calculator)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN public 60; arms/probes/arithmetic.probes.jsonl (dev); arms/probes/arithmetic.heldout.jsonl; the retrieval research agent aligned-en 1010; terse-live EN 213; the evaluation/safety research agent emergency-dev EN 105 / all 300; the evaluation/safety research agent chat-slice EN 66", "judge": "none (deterministic gate)", "n": "60 v0.1 EN + 30 pos/15 neg dev probes + 1,467 non-arithmetic dev messages + 300 emergency-dev + 32 pos/18 neg held-out…", "seeds": "n/a (deterministic); L1 s44+s45", "result": "n/a at L0 (deterministic); L1 component vit-sci-003 rubric +3.33 (1.00 -> 4.33, item ratio 0.200 -> 0.867)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (L0 PASS on A1, A3-A6; A2 FAIL reported separately; L1 AR ADOPT; shipped in cut 3)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [ "F-701" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-003", "code": "CF", "name": "CF", "family": "packs", "question": "Does a deterministic country-facts lookup pack (Wikidata CC0, 197 states) on cut 3 pass the L0 deterministic gates (data accuracy, blind trigger recall/false fires, v0.1+V-exact replay, unit tests, latency) before an L1 judged run vs cut 3 traces/verdicts?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 60 (replay, stub LLM); V-exact 11; venue6 6; CF trigger-set (arms/cf/trigger-set, blind-written, 200 items); the retrieval research agent dev pools 1,328 EN messages (aligned 1010 + terse-l", "judge": "none (deterministic gate)", "n": "data review 30 rows x 5 fields = 150; trigger blind set 200 items (100 positives / 100 lookalikes); dev pool 1,328 EN m…", "seeds": "n/a (deterministic L0; L1 never run)", "result": "data agreement 84 % (bar 98 %)", "result_verified": true, "effect": "data agreement 84 % (bar 98 %)", "ci": null, "mde": "n/a (step-0 gate)", "verdict": "closed", "verdict_detail": "closed: FAIL at L0 (bar 1 data review 84.0% vs >= 98% with currency/plug errors; bar 2 blind trigger recall 80/100 vs >= 90%, false fires 0/100 PASS); bars 3-5 PASS; CF NOT adopted; 0 judge calls", "reopen_if": "none stated (errors: currency 3, plug type 7 + 5 unclear)", "closed_row": "R50", "registry_public": "boar-research:reports/CLOSED.md:79", "date": "2026-10-04", "fact_ids": [ "F-537" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-004", "code": "CUT3-CONFIRM", "name": "CUT3-CONFIRM", "family": "claim-build", "question": "Does cut 3 (cut 2 + VG v3 + AR + SBt, patch 2fbc7001) show no regression vs cut 2 on v0.1 EN and V-exact (identity vs the cut3-vgar L1 arm)?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1; V-exact", "judge": "Claude Opus", "n": "57 judged v0.1 EN items x 2 seeds (5 changed items judged/reused, 52 reused cut 2 verdicts; 4 new calls); V-exact 5 jud…", "seeds": "v0.1 s44, s45; V-exact s42, s43", "result": "gated: min changed-item rubric Δ +0.92 (dng-001), V-exact max drop −0.051; reported (not gated): ratio 0.5656 vs 0.5335, Δ +0.032, CI90 [+0.010, +0.062] (judge.v1-med, cut 2 verdicts reused on 52 items)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "confirmed (no regression) — confirmation of the cut 3 build adopted at QUEUE.md:569", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-005", "code": "CUT3-VGAR-L1", "name": "CUT3-VGAR-L1", "family": "4B-fast-path", "question": "Stacked L1 of cut 2 + VG (verdict guard v1) + AR (arithmetic tools) vs cut 2 on v0.1 EN, with per-component gates on the 4 target items and stack gates (unsafe, cw, paired delta >= 0).", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1", "judge": "Claude Opus", "n": "57 v0.1 EN judged items x 2 seeds; 4 changed items judged (16 calls), 53 control verdicts reused", "seeds": "44, 45", "result": "Δ +0.029, CI90 [+0.007, +0.058] (pooled 0.5335 -> 0.5623, 57 items x 2 seeds, judge.v1-med, changed-items-only reuse)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "AR ADOPT; VG L1 PASS but NOT adopted under this criterion (vg-REVIEW.md V5 FAIL). VG v3 was adopted later through a different gate (VG-S PASS, QUEUE.md:569)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-006", "code": "CUT3DEV", "name": "CUT3DEV (cut 3 build: cut 2 + VG v3 + AR + SBt)", "family": "claim-build", "question": "Build the three 4B-LOSS-ANATOMY arms (VG, AR, SC) as commits on cut 2 and assemble the adoptable cut 3, measured by the stacked L1 vs cut 2 on v0.1 EN.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN; V-exact (confirmation)", "judge": "Claude Opus", "n": "57 judged v0.1 EN items (changed items only judged)", "seeds": "s44+s45", "result": "stack L1 (VG v1 + AR vs cut 2) Δ +0.029, CI90 [+0.007, +0.058], 0.5335 → 0.5623; confirmation (VG v3 + AR + SBt) Δ +0.032, CI90 [+0.010, +0.062], 0.5335 → 0.5656", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (CUT 3 ADOPTED after VG-S, CONFIRMED with no regression)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-007", "code": "CV", "name": "CV", "family": "4B-fast-path", "question": "Does a computed-verdict path (4B extracts facts in a constrained format, engine computes the verdict deterministically) on cut 3 raise verdict accuracy by >= 10 pp over cut 3 with <= 2 new wrong verdicts, 0 control fires and <= +2 s added time, on DEV and then on a fresh blind verdict gate set?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "vg-probe-q + vg-probe-q2 DEV (210); CV gate set (arms/cv/gate-set, 120); the retrieval research agent EN dev messages 1,328; v0.1 EN (routing only); V-exact (routing only)", "judge": "none (deterministic gate)", "n": "DEV 155 verdict items + 55 controls (145 scored); gate set 120 (90 verdict + 30 controls, 83 scored)", "seeds": "s44", "result": "gate gains +6 to +26.5 pp but regressions or L1 Δ ≤ 0", "result_verified": true, "effect": "gate gains +6 to +26.5 pp but regressions or L1 Δ ≤ 0", "ci": null, "mde": "n/a (deterministic gates)", "verdict": "closed", "verdict_detail": "closed: KILLED. L0 FAIL (6 regressions, bar <= 2) → revision 1 → ship gate FAIL (3 regressions, bar <= 2) → stop rule (QUEUE.md:596). The computed-verdict family (CV/CVS/CVH/CVO) was closed later (QUEUE.md:604, 618)", "reopen_if": "QUEUE l.604: verdict-only override where the answer disagrees with a sane computed verdict (= CVO, R56, also killed)", "closed_row": "R53", "registry_public": "boar-research:reports/CLOSED.md:82", "date": "2026-10-04", "fact_ids": [ "F-540", "F-245", "F-262" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-008", "code": "K3-ON-CUT2", "name": "K3-ON-CUT2", "family": "4B-fast-path", "question": "Can the 4B fast tier retrieve 3 passages instead of 4 (quickBudget() non-lookup retrieveK 4 -> 3, speed lever K3) without measurable quality loss, vs ship cut 2 (cff2f2bc) overnight traces?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1", "judge": "none (deterministic gate)", "n": "60 v0.1 EN items x 2 seeds (100 paired item x seed on the 4B answer path); 0 judge calls", "seeds": "44, 45", "result": "n/a (quality not measured); gate 5 prompt-token saving 19.3 (458.9 -> 439.7) < 100 bar", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed (NO-ADOPT at deterministic gate 5; judge skipped by frozen cost rule; gates 1, 2, 4 not measured)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-009", "code": "SC-L0", "name": "SC-L0 (SC v1 trigger: water-safety card + snakebite completion)", "family": "4B-fast-path", "question": "Does a deterministic EN water-safety card trigger (plus unchanged snakebite trigger) fire on dng-005 and drinking-water paraphrases with 0 false fires and >= 80 % held-out recall, vs cut 2 (no water card)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN public (dng-005); arms/probes/water-safety.probes.jsonl (dev); arms/probes/water-safety.heldout.jsonl; EN dev pools 1,328; CO lookalike dev/heldout/heldout3; chat-slice EN 66; the evaluation/safety research agent emergency-dev emergencies 180", "judge": "none (deterministic gate)", "n": "held-out 25 pos / 25 neg; dev pools 1,703; 22/18 dev paraphrases; 180 emergencies; 2,043 snakebite diff messages", "seeds": "n/a", "result": "water recall 72 %/84 %/72 %; false fires 0/5/2", "result_verified": true, "effect": "water recall 72 %/84 %/72 %; false fires 0/5/2", "ci": null, "mde": "n/a (gate)", "verdict": "closed", "verdict_detail": "FAIL (S5 72 % < 80 %); superseded by v2/v3; water card and trigger PARKED", "reopen_if": "see R59–R60", "closed_row": "R58", "registry_public": "boar-research:reports/CLOSED.md:87", "date": "2026-10-04", "fact_ids": [ "F-423" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-010", "code": "SC-S6", "name": "SC-S6 / SBt (reviewed snakebite card text)", "family": "4B-fast-path", "question": "Does the independent Opus reviewer approve the EN water-safety and snakebite card texts (criterion S6), and does the reviewed snakebite text ship on cut 2's unchanged trigger?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "CDC/EPA/WHO/NIOSH sources fetched 2026-10-04", "judge": "Claude Opus, independent reviewer agent", "n": "2 cards", "seeds": "n/a", "result": "dng-001 rubric +0.92 (2.67 -> 3.58), cut 3 confirmation, judge.v1-med, s44+s45", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (snakebite text only, SBt in cut 3; water text approved but never shipped)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-011", "code": "SC-V2-S5PRIME", "name": "SC-V2-S5PRIME (sc-fresh3)", "family": "packs", "question": "Does the frozen SC v2 safety-card trigger (water / snakebite cards, commit 02406703) reach S5′ on fresh blind set #3 (100 items: water 25, snakebite 25, none 50): recall ≥ 80 % each and false fires ≤ 1 per 25 none items?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "sc-fresh3", "judge": "none (deterministic gate)", "n": "100 (water 25, snakebite 25, none 50)", "seeds": "n/a (deterministic trigger, scored once)", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed — S5′ FAIL (false fires 5/50 vs ≤ 1 per 25); follow-up SC v3 S5″ also FAIL, SC trigger parked", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-012", "code": "SC-V3-S5SECOND", "name": "SC-V3-S5SECOND (SC v3 trigger S5″ on sc-fresh4)", "family": "safety-cards / routing trigger (4B fast path)", "question": "Does the SC v3 safety-card trigger (water + snakebite fast path, with negation / non-human victim / present-need guards) reach >=80% recall per card with <=1 false fire per 25 none-items on the fresh blind-written set #4, vs the frozen bars identical to S5′ (SC v2 failed on set #3)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "sc-fresh4 (SC fresh set #4, items.jsonl sha256 653b1a99…e125)", "judge": "none (deterministic gate)", "n": "100 items (25 water / 25 snakebite / 50 none)", "seeds": "n/a (deterministic trigger, scored once)", "result": "n/a (deterministic gate: water 18/25 (72 %) FAIL, snakebite 20/25 (80 %) PASS, false fires 2 on 50 PASS)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed (S5″ FAIL: water recall 18/25 (72 %) < 80 %; SC water card + trigger changes parked for the bounty)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-013", "code": "TF2-ON-CUT2", "name": "TF2-ON-CUT2", "family": "retrieval", "question": "Does the retrieval research agent's tf2 (cut 2 + exp-39 travel/food routing fixes + exp-40 crypto-kit gate) go into the next ship cut, vs ship cut 2?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1; V-exact", "judge": "Claude Opus", "n": "G2: 60 v0.1 EN items, 1 seed; G3: 5 V-exact judge items x 2 seeds x 2 orders = 20 judge calls", "seeds": "G2 s44; G3 s42, s43", "result": "vx-001 Japan −0.018 (0.356 -> 0.338 (0.276, 0.400)); bar >= 0.356 FAIL", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed (NO-GO: G3 Japan bar failed; tf2 stays out of the ship cut)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-014", "code": "VG-L0", "name": "VG-L0 (verdict guard v1)", "family": "4B-fast-path", "question": "Does a regex verdict guard that detects and template-repairs Yes/No verdicts contradicting the answer's own figures (date/compare/rank) flag the 3 v0.1 targets without touching guard items, vs cut 2 (no guard)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN traces; the evaluation/safety research agent chat-slice EN; EN dev pools 1,328", "judge": "none (deterministic gate)", "n": "12 trace runs x 60 (CB 57) v0.1 EN; 506 chat answers; 1,328 dev questions; 32 repair pairs", "seeds": "cut 2 s44-s47, A s42/s43, E0 s42-s44, RV2 s42/s43, CB x1", "result": "n/a (L0); stacked L1 component deltas lan-ref-003 +2.67, vit-sci-002 +1.08, lan-cmp-003 +1.00 (judge.v1-med, s44+s45)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "L0 FAIL by the letter (V2) with a disclosed exception, plus V5 FAIL. The VG line was later adopted as v3 in cut 3.", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-015", "code": "VG-PROBE-Q", "name": "VG-PROBE-Q", "family": "tooling", "question": "Fresh blind-written probe set (90 Qs: compare2 25 / threshold 20 / rank 20 / control 25) used as the VG-S ship gate for verdict guard VG v3 on cut 2 answers; holds no verdict of its own.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "vg-probe-q", "judge": "none (deterministic gate)", "n": "90", "seeds": "n/a (answers generated with cut 2, s44 per QUEUE l.568)", "result": "n/a (VG-S: 1 fire / 90 (vgq-013, BETTER), 0 control fires, 90/90 valid)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling (probe set; VG-S PASS on it; DEV after VG-S)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-016", "code": "VG-PROBE-Q2", "name": "VG-PROBE-Q2", "family": "tooling", "question": "Second fresh blind-written probe set (120 Qs: threshold / superlative / compare2 / control 30 each) used as the VG-S2 ship gate for VG v4; holds no verdict of its own.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "vg-probe-q2", "judge": "none (deterministic gate)", "n": "120", "seeds": "n/a (answers generated with cut 2, s44 per QUEUE l.572)", "result": "n/a (VG-S2: 2 fires, vgq2-002 WRONG + vgq2-119 BETTER, 0 control, 120/120 valid)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling (probe set; VG-S2 FAIL on it; burnt)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "agent-lab", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-017", "code": "VG-S", "name": "VG-S (VG v3 ship gate)", "family": "4B-fast-path", "question": "On a fresh blind-written 90-question probe set answered once by cut 2, does VG v3 produce 0 WORSE/WRONG fires and 0 control fires (ship gate for cut 3)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "arms/vg-probe-q/items.jsonl (blind-written probe set)", "judge": "Claude Opus, independent reviewer agent", "n": "90 (compare2 25 / threshold 20 / rank 20 / control 25); 81 checked", "seeds": "s44 (single generation)", "result": "n/a for the gate; cut 3 stack L1 Δ +0.029, CI90 [+0.007, +0.058] (VG v1 + AR); confirmation +0.032, CI90 [+0.010, +0.062]", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (VG-S PASS → cut 3 adopted with VG v3)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-018", "code": "VG-V5prime", "name": "VG-V5prime (VG v2 re-review + v3 equivalence)", "family": "4B-fast-path", "question": "After the v2 fixes (list-consistent rank, quantity/unit-matched compare), are all repaired pairs BETTER with 0 WORSE/WRONG (V5′), and is v3 (F1–F4 abstain rules) byte-equivalent to v2 on stored answers?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN stored traces (L0 runs + cut3-vgar L1 control copies); chat dev", "judge": "Claude Opus, independent reviewer agent", "n": "36 repair pairs; 2,326 stored EN model answers", "seeds": "n/a", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "intermediate gate PASS (V5′ 36/36 BETTER); not a ship clearance; superseded by v3, which shipped after VG-S", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-04", "fact_ids": [], "chapter": "walls", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-019", "code": "VG4-S2", "name": "VG4-S2 (VG v4 threshold/superlative ship gate)", "family": "4B-fast-path", "question": "Does VG v4 (adds numeric-threshold and superlative yes/no checks) pass a ship gate of 0 WORSE/WRONG fires and 0 control fires on fresh probe set #2, vs cut 3 with VG v3?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "arms/vg-probe-q2/items.jsonl (blind-written probe set #2); arms/vg-probe-q (DEV for v4); EN dev pools 1,328; 2,426 stored EN answers", "judge": "Claude Opus, independent reviewer agent", "n": "120 (threshold 30 / superlative 30 / compare2 30 / control 30); 103 checked", "seeds": "s44 (single generation)", "result": "1 WRONG of 2 fires; regex catches 1/17 self-contradictions", "result_verified": true, "effect": "1 WRONG of 2 fires; regex catches 1/17 self-contradictions", "ci": null, "mde": "n/a (reviewer gate)", "verdict": "adopted", "verdict_detail": "closed (VG-S2 FAIL: 1 WRONG fire; VG v4 not adopted; VG line CLOSED)", "reopen_if": "none: model-side self-contradiction (QUEUE l.577)", "closed_row": "R57", "registry_public": "boar-research:reports/CLOSED.md:86", "date": "2026-10-04", "fact_ids": [ "F-541" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-020", "code": "SC", "name": "Safety-card trigger v2", "family": "4B", "question": "Safety-card trigger v2", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": null }, "set": "100 items (25 water / 25 snakebite / 50 none), scored once, no judge", "judge": null, "n": null, "seeds": null, "result": "water recall 72 %/84 %/72 %; false fires 0/5/2", "result_verified": true, "effect": "water recall 72 %/84 %/72 %; false fires 0/5/2", "ci": null, "mde": "n/a (gate)", "verdict": "closed", "verdict_detail": "closed (gate FAIL; set #3 burned)", "reopen_if": "see R60", "closed_row": "R59", "registry_public": "boar-research:reports/CLOSED.md:88", "date": "2026-10-04", "fact_ids": [ "F-423", "F-406" ], "chapter": "builds", "source": "closed registry row R59 + verified fact" }, { "id": "X-021", "code": "SC", "name": "Safety-card trigger v3", "family": "4B", "question": "Safety-card trigger v3", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": null }, "set": "100 items (25/25/50), scored once, no judge", "judge": null, "n": null, "seeds": null, "result": "water recall 72 %/84 %/72 %; false fires 0/5/2", "result_verified": true, "effect": "water recall 72 %/84 %/72 %; false fires 0/5/2", "ci": null, "mde": "n/a (gate)", "verdict": "not-detected", "verdict_detail": "closed (gate FAIL); water card + SC trigger PARKED for the bounty; the snakebite text (SBt) shipped separately in cut 3", "reopen_if": "none before the bounty (QUEUE l.562)", "closed_row": "R60", "registry_public": "boar-research:reports/CLOSED.md:89", "date": "2026-10-04", "fact_ids": [ "F-423", "F-406" ], "chapter": "builds", "source": "closed registry row R60 + verified fact" }, { "id": "X-022", "code": "CRYPTOSEL", "name": "CRYPTOSEL", "family": "retrieval (selection / crypto pack)", "question": "On cut 3, does crypto-pack spec selection (prefer chunks naming the EIP number/constant, keep constants tables whole, same budget) put the gold fact in context for vit-cry-002/003/009 and lift their rubric by >= +0.5 at L1 vs cut 3, with non-crypto questions untouched?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1 EN; V-exact (L0 only)", "judge": "Claude Opus", "n": "L0: 71 items (v0.1 EN 60 + V-exact 11); L1: 57 judged items x 2 seeds, 4 changed items judged", "seeds": "44, 45", "result": "+0.0033 (CI90 −0.005…+0.014)", "result_verified": true, "effect": "+0.0033 (CI90 −0.005…+0.014)", "ci": null, "mde": "≈ 0.036 (v0.1, 2 seeds)", "verdict": "closed", "verdict_detail": "closed (L1 FAIL on target gate, not adopted; stop rule: no revision after L1)", "reopen_if": "none stated", "closed_row": "R52", "registry_public": "boar-research:reports/CLOSED.md:81", "date": "2026-10-05", "fact_ids": [ "F-539", "F-1096" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-023", "code": "CUT4-INSTANT-FRAG", "name": "CUT4-INSTANT-FRAG", "family": "4B-fast-path (instant tier fix, claim-build cut 4)", "question": "Does blocking list/heading fragments and figurative 'capital' matches in the instant-final decision stop off-topic instant answers (France warm-up -> Lyon line) without losing correct instant answers or changing v0.1 quality vs cut 3?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN; V-exact; 1,328 the retrieval research agent EN dev messages; cut4 lookup-set (40, coordinator-frozen)", "judge": "none (deterministic gate)", "n": "L0: 71 (v0.1 EN 60 + V-exact 11) + 1 warm-up + 1,328 dev messages + 40 lookup set; L1: v0.1 EN 60 x 2 seeds", "seeds": "44, 45 (L1); stub replay at L0", "result": "n/a (0 changed items vs same-day cut 3 on v0.1 EN s44+s45; judge skipped)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (L0 PASS, L1 PASS vacuously with empty changed set; shipped in cut 4, later extended by INSTANT-SUBJ)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-05", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-024", "code": "CUT4-KVFIX", "name": "CUT4-KVFIX", "family": "device (latency fix, claim-build cut 4)", "question": "Does sanitizing the KV session key and creating its directory fix the A15 saveSession failure (gate/answer prefixes re-prefilled every question) without changing desktop answers vs cut 3?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN", "judge": "none (deterministic gate)", "n": "v0.1 EN 60 x 2 seeds (desktop identity); 4 unit tests", "seeds": "44, 45", "result": "n/a (identity gate: 60/60 identical per seed vs same-day cut 3, s44 and s45; judge skipped)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (shipped in cut 4 on gates (1)+(2) per criterion; device gate (3) still open, never reported)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-05", "fact_ids": [], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-025", "code": "CVH", "name": "CVH", "family": "4B-fast-path (computed-verdict line)", "question": "Does passing the CVS computed verdict + facts as a hint into the normal cut 3 answer path (keeping prose and citations) avoid the CVS judged-quality loss on v0.1 EN L1 vs cut 3 (Step 1), before a fresh blind verdict gate #4 (Step 2)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN", "judge": "Claude Opus", "n": "v0.1 EN 60 answers/seed, 57-item paired Δ; 4 changed items judged (16 calls)", "seeds": "s44 + s45", "result": "gate gains +6 to +26.5 pp but regressions or L1 Δ ≤ 0", "result_verified": true, "effect": "gate gains +6 to +26.5 pp but regressions or L1 Δ ≤ 0", "ci": null, "mde": "≈ 0.036", "verdict": "closed", "verdict_detail": "closed (Step 1 v0.1 EN L1 FAIL on 3 of 5 gates; Step 2 never started by the stop rule)", "reopen_if": "as R53", "closed_row": "R55", "registry_public": "boar-research:reports/CLOSED.md:84", "date": "2026-10-05", "fact_ids": [ "F-540", "F-245", "F-262" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-026", "code": "CVO", "name": "CVO", "family": "4B-fast-path (computed-verdict line)", "question": "Does overriding only the verdict sentence of cut 3's answer when it disagrees with a sane computed verdict (prose otherwise byte-identical) pass v0.1 EN L1 vs cut 3 and raise verdict accuracy on a fresh blind gate #4 by >= +8 pp?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1 EN; CVO gate set #4 (gate4-set 0e1781e9…, key d51bcd62…)", "judge": "Claude Opus", "n": "L1: v0.1 EN 60/seed, 57 paired; gate #4: 120 items, 84 scored (6 ambiguous)", "seeds": "L1: s44 + s45; gate #4: not stated in arm files", "result": "gate gains +6 to +26.5 pp but regressions or L1 Δ ≤ 0", "result_verified": true, "effect": "gate gains +6 to +26.5 pp but regressions or L1 Δ ≤ 0", "ci": null, "mde": "n/a (gate)", "verdict": "closed", "verdict_detail": "closed (Step 1 L1 PASS vs the same-day control; Step 2 fresh blind gate #4 FAIL, +6.0 pp < +8 bar; killed by the stop rule)", "reopen_if": "none for the bounty (QUEUE l.604)", "closed_row": "R56", "registry_public": "boar-research:reports/CLOSED.md:85", "date": "2026-10-05", "fact_ids": [ "F-540", "F-245", "F-262" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-027", "code": "CVS", "name": "CVS", "family": "4B-fast-path (computed-verdict line)", "question": "Does CV rev1 (computed verdict for compare2 + threshold questions) plus a deterministic fact-sanity guard (per-class plausibility ranges, same-class 100x pair rule, abstain to normal path) beat cut 3 on a fresh blind verdict gate and survive v0.1 EN L1 vs cut 3?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "CVS gate set #3 (gate3-set 6af3b0d1…, key 6c3757df…); v0.1 EN; burned CV gate set (descriptive); DEV probes (cv-dev-cv-s44)", "judge": "none (deterministic gate)", "n": "gate #3: 120 items (83 scored, 7 ambiguous key lines); L1: v0.1 EN 60 answers/seed, 57-item paired Δ, 4 changed items j…", "seeds": "gate #3: s44 (1 seed); L1: s44 + s45", "result": "gate gains +6 to +26.5 pp but regressions or L1 Δ ≤ 0", "result_verified": true, "effect": "gate gains +6 to +26.5 pp but regressions or L1 Δ ≤ 0", "ci": null, "mde": "≈ 0.036", "verdict": "closed", "verdict_detail": "closed (gate #3 PASS, v0.1 EN L1 FAIL on 2 gates; killed by the no-revision stop rule)", "reopen_if": "as R53", "closed_row": "R54", "registry_public": "boar-research:reports/CLOSED.md:83", "date": "2026-10-05", "fact_ids": [ "F-540", "F-142", "F-245", "F-262" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-028", "code": "G4E4B", "name": "G4E4B", "family": "4B-fast-path", "question": "Closed-book screen: does Gemma 4 E4B QAT Q4_0 beat the Qwen3-4B-Instruct-2507 baseline (0.487) by >=0.03 on v0.1 EN 57, with decode >=0.8x, RSS <=7.0 GiB, 0 empties and no token leaks?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 57 (closed-book)", "judge": "Claude Opus", "n": "57 items, 114 judge calls", "seeds": "1 (arm seed 44; control seed 42)", "result": "0.437/0.495/0.488/0.457 vs 0.487", "result_verified": true, "effect": "0.437/0.495/0.488/0.457 vs 0.487", "ci": null, "mde": "0.055", "verdict": "not-detected", "verdict_detail": "closed (screen FAIL on quality only; stop rule ends G4E4B; G4ENG later declared post-screen)", "reopen_if": "a <=9B model with published closed-book SimpleQA / AA-Omniscience accuracy clearly above ~20 %", "closed_row": "R1", "registry_public": "boar-research:reports/CLOSED.md:16", "date": "2026-10-05", "fact_ids": [ "F-512" ], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-029", "code": "G4ENG", "name": "G4ENG", "family": "4B-fast-path", "question": "Does swapping the Android 12 GB reader to Gemma 4 E4B (q4_0) inside the full cut 3 engine beat same-day cut 3 (Qwen3-4B) on v0.1 EN?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN", "judge": "Claude Opus", "n": "60 items (57 judged) x 2 seeds; 456 judge calls", "seeds": "s44, s45", "result": "Δ −0.0385 (CI90 −0.073…−0.006); 1.49× slower", "result_verified": true, "effect": "Δ −0.0385 (CI90 −0.073…−0.006); 1.49× slower", "ci": null, "mde": "~0.036 (QUEUE l.103)", "verdict": "closed", "verdict_detail": "closed (FAIL on 5 gates, not adopted, G4ENG ended)", "reopen_if": "none stated (stop rule)", "closed_row": "R2", "registry_public": "boar-research:reports/CLOSED.md:17", "date": "2026-10-05", "fact_ids": [ "F-420" ], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-030", "code": "INSTSUBJ", "name": "INSTSUBJ", "family": "4B-fast-path (instant tier)", "question": "Cut 4 + INSTANT-SUBJ (an instant extractive snippet is final only when the question's entity is the subject/possessor of the answered slot) vs cut 4: L0 stub replay with 0 wrong instant-final answers and <= 1 correct instant lost on lookup set #2, 0 correct instants lost on v0.1 EN + V-exact.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "lookup set #2 (arms/instsubj/lookup2); v0.1 EN public 60; V-exact 11; the retrieval research agent EN dev messages 1,328", "judge": "none (deterministic gate)", "n": "lookup set #2: 55 items (40 lookup, 15 control); v0.1 EN 60 + V-exact 11 + France warm-up = 72 requests (66 unique ids)…", "seeds": "n/a (stub-LLM replay)", "result": "n/a (0 tier changes, 0 correct instants lost on v0.1 EN + V-exact and 1,328 dev; lookup2 instant-final 0 in both arms)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted into cut 4 (L0 PASS, uninformative gate; QUEUE.md:647)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-05", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-031", "code": "KVQ8", "name": "KVQ8", "family": "4B-fast-path", "question": "Quality non-inferiority of q8_0 KV cache vs f16 for the 4B reader on cut 4 (+instsubj), v0.1 EN two seeds, interleaved same-session control.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1", "judge": "Claude Opus", "n": "57 judged items x 2 seeds (60 items run per seed); 180 Opus calls on changed items", "seeds": "2 (s44, s45)", "result": "−0.0077, two-level bootstrap CI90 [−0.0232, +0.0062] (pooled s44+s45; s44 −0.0011, s45 −0.0142)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "PASS all gates → ADOPT as a device-day candidate (device speed/swap untested; not in any cut)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-05", "fact_ids": [], "chapter": "agent-lab", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-032", "code": "PROMPTDIET", "name": "PROMPTDIET", "family": "4B-fast-path", "question": "Can the fixed prompt text on the 4B normal answer path (cut 3) be cut by >= 150 mean tokens without dropping or weakening any rule (L0), before an L1 non-inferiority test vs same-day cut 3 control?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1", "judge": "none (deterministic gate)", "n": "100 v0.1 EN cut3-confirm traces (s44 + s45) for the L0 token count; no L1 run", "seeds": "s44 + s45 traces (token measurement only)", "result": "quality −0.016; confident-wrong 5 → 9", "result_verified": true, "effect": "quality −0.016; confident-wrong 5 → 9", "ci": null, "mde": "~0.036", "verdict": "closed", "verdict_detail": "closed: L0 FAIL after the one allowed revision (−125.72 vs bar −150); never ran L1 under PROMPTDIET; the faithful diet was later tested inside SPEEDPACK L1 (FAIL) and dropped from device-day candidates (QUEUE l.660)", "reopen_if": "a new device-day latency experiment (patch 9904d4e1 kept); see QUEUE l.614, l.660", "closed_row": "R7", "registry_public": "boar-research:reports/CLOSED.md:22", "date": "2026-10-05", "fact_ids": [ "F-516", "F-1097" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-033", "code": "Q35", "name": "Q35", "family": "4B-fast-path", "question": "Screen of Qwen3.5-4B (Q4_K_M, non-thinking) as the 12 GB default reader vs Qwen3-4B-Instruct-2507 closed-book baseline 0.487 on v0.1 EN (GO iff ratio >= 0.517, decode >= 0.8x, RSS <= +0.5 GiB, 0 empty, no think leaks).", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1", "judge": "Claude Opus", "n": "57 items, 114 judge calls (both orders)", "seeds": "seed 44 (control seed 42)", "result": "0.437/0.495/0.488/0.457 vs 0.487", "result_verified": true, "effect": "0.437/0.495/0.488/0.457 vs 0.487", "ci": null, "mde": "0.055", "verdict": "not-detected", "verdict_detail": "closed: screen FAIL on the quality bar (closed-book ratio 0.437 < 0.517; below the 0.487 baseline); speed/RSS/empty/leak bars PASS; no full-engine L1 (stop rule)", "reopen_if": "a <=9B model with published closed-book SimpleQA / AA-Omniscience accuracy clearly above ~20 %", "closed_row": "R1", "registry_public": "boar-research:reports/CLOSED.md:16", "date": "2026-10-05", "fact_ids": [ "F-512" ], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-034", "code": "TRAVELSEL", "name": "TRAVELSEL", "family": "retrieval", "question": "Does preferring the matching Wikivoyage practical section (Get in / Get around / Stay safe / Electricity / Money...) of a by-name destination article put in-pack gold travel facts into the cut 3 context (L0 stub replay), vs cut 3 base contexts?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1; V-exact", "judge": "none (deterministic gate)", "n": "3 target items (gold-fact gate); 71 items replayed (v0.1 EN 60 + V-exact 11)", "seeds": "none (deterministic stub-LLM replay); L1 (s44+s45) never run", "result": "1/3 (bar 2)", "result_verified": true, "effect": "1/3 (bar 2)", "ci": null, "mde": "n/a (gate)", "verdict": "closed", "verdict_detail": "closed (L0 FAIL after the one allowed revision; QUEUE calls it 'killed'); L1 never run", "reopen_if": "none stated (\"fails on structure, not on tuning\")", "closed_row": "R64", "registry_public": "boar-research:reports/CLOSED.md:93", "date": "2026-10-05", "fact_ids": [ "F-1125", "F-408" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-035", "code": "CHATGATE", "name": "CHATGATE", "family": "retrieval", "question": "Descriptive arm: does the route gate's CONVERSATION branch on cut 5 (ship profile, gate on) route the 192 dev chat-slice messages correctly vs their expected_search labels (missed search, false search, pasted passages)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "dev chat slices v0/v1/v2 (192 conversations)", "judge": "none (deterministic gate)", "n": "192 conversations (69 expected yes / 102 no / 21 either)", "seeds": "1 (s44)", "result": "n/a (descriptive; no judge)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "closed (descriptive arm): G1 FAIL as frozen (39/69 = 56.5%, bar <= 5%), G2 PASS (1/102 = 1.0%), G3 PASS (0/0/0). Reading: the gate is fine (true gate miss 1/69, v1-055); the G1 shortfall is kit coverage; no gate-fix arm", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-036", "code": "COMPLETE", "name": "COMPLETE", "family": "deep-tier", "question": "Does a completeness instruction (120-220 words, max tokens 700) on the deep35 reader with gold R450 contexts raise the judge ratio vs the identical DEEPREAD run (short-answer prompt) without losing correctness?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 57", "judge": "Claude Opus", "n": "57 (v0.1 EN)", "seeds": "1 (seed 42)", "result": "−0.014 (−0.055…+0.024); confident-wrong 7 → 15", "result_verified": true, "effect": "−0.014 (−0.055…+0.024); confident-wrong 7 → 15", "ci": "−0.055…+0.024", "mde": "0.055", "verdict": "closed", "verdict_detail": "closed: NO by the frozen rule (ratio 0.620 < 0.700; correctness 3.10 vs >= 3.80; confident-wrong 7 → 15 vs allowed +2)", "reopen_if": "none stated; DEEPCOVER (closed NO 10-12, R67; key-point coverage, BACKLOG B2) was the only distinct variant", "closed_row": "R8", "registry_public": "boar-research:reports/CLOSED.md:23", "date": "2026-10-06", "fact_ids": [ "F-517", "F-653" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-037", "code": "CRYPTOTRUTH", "name": "CRYPTOTRUTH", "family": "4B-fast-path (honesty / deterministic EIP card)", "question": "Does cut 6 + NOREC (fixed honest 'couldn't find X offline' line replacing unsupported 4B denials) + EIPABS (verbatim EIP card bypassing the 4B) raise v0.1 quality and cut confident-wrong vs same-day cut 6, while giving honest not-found answers on 3 frozen fictional probes?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN; 192 dev chat messages; 3 frozen fictional probes", "judge": "Claude Opus", "n": "57 judged items x 2 seeds (5 changed item-seeds judged); 192 dev chat messages; 3 fictional probes", "seeds": "44, 45", "result": "+0.0092, CI90 [-0.0034, +0.0244] (ctl 0.5594 -> arm 0.5686; s44 +0.0095, s45 +0.0089)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "parked (NOT ADOPTED as frozen: G2 FAIL on 1 of 3 fictional probes; G1 and G3 pass; QUEUE l.702 'parked for after the claim (PR phase), with a broadened phrase list as a new freeze')", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [ "F-411" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-038", "code": "CUT5-LITE", "name": "CUT5-LITE", "family": "claim-build", "question": "Is cut 4 + only the second KV prefix (.nosrc) and the tightened route-gate skip (perf arms removed) build-clean, answer-identical to a same-day cut 4 control, safe on the 464 dev messages, and not slower on desktop retrieval (<= 1.03x)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN; 464 dev messages (same as CUT5-SPEED G4)", "judge": "none (deterministic gate)", "n": "120 (v0.1 EN 60 x 2 seeds); G4 464 dev messages", "seeds": "s44, s45", "result": "retrieval wall arm/control pooled 0.961 (14600 -> 14031 ms; s44 0.989, s45 0.934), bar <= 1.03; no CI; within ~5% control noise", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (cut 5 = device-day candidate; later the base of cut 6 = cut 5 + DETMAX, QUEUE.md:694, and so of cut 7)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-039", "code": "CUT5-SPEED", "name": "CUT5-SPEED", "family": "claim-build", "question": "Does cut 4 + three latency-only changes (the native/deep-tier research agent retrieval perf arms: native zstd/vector cache/sqlBatch; second KV prefix for the no-source system message; route-gate skip for lookup/research/compare) keep answers identical to a same-day cut 4 control while cutting desktop retrieval wall time to <= 0.8x and never skipping the gate on small talk/BOAR/emergency messages?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN; chat-slice v0/v1/v2 (dev); bench v0.1 public EN+PT; coordinator probes vg; water-safety probes + heldout; arithmetic probes; sonar travel/food probes", "judge": "none (deterministic gate)", "n": "120 (v0.1 EN 60 items x 2 seeds; 57 items/run with retrieval timing); G4 464 dev messages", "seeds": "s44, s45", "result": "retrieval wall arm/control pooled 1.071 (14457 -> 15478 ms; s44 1.082, s45 1.059), bar <= 0.800; no CI; control run-to-run noise 4.7%", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed: NOT ADOPTED as cut 5 (G3 FAIL 1.071x vs bar <= 0.800; G1/G2/G4 PASS); perf arms parked for a device ABBA (QUEUE.md:699 A/B 0.445x confounded); components 2+3 adopted via CUT5-LITE", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [ "F-809", "F-818" ], "chapter": "handoff", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-040", "code": "DEEPREAD", "name": "DEEPREAD", "family": "deep-tier", "question": "Does the deep model (Qwen3.6-35B-A3B UD-Q2_K_XL) convert gold retrieved passages (R450 oracle contexts) better than the 4B (Qwen3-4B-Instruct-2507) reading the identical contexts, on v0.1 EN 57?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 57", "judge": "Claude Opus", "n": "57", "seeds": "1 (seed 42)", "result": "+0.029 (−0.004…+0.063)", "result_verified": true, "effect": "+0.029 (−0.004…+0.063)", "ci": "−0.004…+0.063", "mde": "0.055", "verdict": "not-detected", "verdict_detail": "INCONCLUSIVE (frozen band 0.606 ≤ x < 0.660; deep35 R450 0.6345); CLOSED R15 records it as 'not detected at MDE 0.055'", "reopen_if": "none stated", "closed_row": "R15", "registry_public": "boar-research:reports/CLOSED.md:34", "date": "2026-10-06", "fact_ids": [ "F-522" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-041", "code": "DETFU", "name": "DETFU", "family": "device", "question": "Diagnostic: with the real learned emergency detector (cut 5), do emergencies split across turns (S1/S2) or preceded by a chit-chat turn (S3 control) lose the emergency/danger route of their standalone form?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/detfu/DETFU-criterion.md" }, "set": "dev emergency probes (emergency-dev 49, v0.1 danger 5, sonar CO 7)", "judge": "none (deterministic gate)", "n": "61 probes (60 with standalone emergency route)", "seeds": "n/a (deterministic routing, generation stubbed)", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed (diagnostic): decision rule met through the S3 clause only (S1/S2 0/60; S3 1/60 mismatch). Led to DETMAX", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-042", "code": "DETMAX", "name": "DETMAX", "family": "device", "question": "Does scoring the emergency detector on both the history window and the current turn alone (fire if either ≥ threshold) remove history dilution without changing single-turn behaviour or adding false fires, vs cut 5?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/detfu/DETMAX-criterion.md" }, "set": "v0.1 EN s44; DETFU dev emergency probes + 3 held-out prefixes; 192 dev chat messages", "judge": "none (deterministic gate)", "n": "60 probes × 4 prefixes (240 forms); v0.1 EN 60; 192 dev chat conversations (231 user turns)", "seeds": "v0.1 s44", "result": "n/a (G3: 0/240 arm mismatches vs control 7/240; G4: 1 added route in 192 dev chat conversations)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (ADOPT, all five gates pass) → cut 6 = cut 5 + DETMAX", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [ "F-328", "F-345", "F-701" ], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-043", "code": "FACTS", "name": "FACTS", "family": "4B-fast-path", "question": "Does a deterministic Wikidata structured-facts instant tier (r1 prototype on cut 6) cover enough v0.1 items (step 1) and pass deterministic gates G1-G5 (precheck, false fires, retrieval identity, latency) vs cut 6?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/facts/FACTS-criterion.md" }, "set": "v0.1; V-exact 11 (bench-vslice, dev); dev chat v0-v2 (192); safety track (9 EN + 3 PT)", "judge": "none (deterministic gate)", "n": "v0.1 public 60 + V-exact 11 (step 1); 8 targets; 52 other v0.1, 12 safety, 192 dev chat (G3)", "seeds": "deterministic (replay vs ctl-v01-s44)", "result": "recall 70 % < 75 % bar", "result_verified": true, "effect": "recall 70 % < 75 % bar", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "GO by the letter (step 1 at 8/60, G2-G5 pass); r1 superseded by r2 the same day", "reopen_if": "recall axis only: see FACTS-RECALL.md and LCRELAX-PROBE.md", "closed_row": "R9", "registry_public": "boar-research:reports/CLOSED.md:24", "date": "2026-10-06", "fact_ids": [ "F-421", "F-046", "F-339", "F-343", "F-344", "F-346", "F-347" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-044", "code": "FACTS-BLIND", "name": "FACTS-BLIND", "family": "4B-fast-path", "question": "Does FACTS r3 (the five FACTS-RECALL fixes) hold recall and precision on a second blind recall set (set 2), and what does it score on the now-dev set 1?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "recall-set-2.jsonl (FACTS blind set 2); recall-set.jsonl (dev)", "judge": "none (deterministic gate)", "n": "60 (40 fact + 20 near-miss) on set 2; 60 on set 1 (dev)", "seeds": "deterministic", "result": "recall 70 % < 75 % bar", "result_verified": true, "effect": "recall 70 % < 75 % bar", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed (direction confirmed on precision; recall open; set 2 dev after r3)", "reopen_if": "recall axis only: see FACTS-RECALL.md and LCRELAX-PROBE.md", "closed_row": "R9", "registry_public": "boar-research:reports/CLOSED.md:24", "date": "2026-10-06", "fact_ids": [ "F-421" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-045", "code": "FACTS-JUDGE", "name": "FACTS-JUDGE", "family": "4B-fast-path", "question": "Do the FACTS r2 deterministic answers on the 8 fired v0.1 items score at least as well as the cut 6 control (spctl-v01-s44) under judge.v1-med?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/facts/FACTS-judge-criterion.md" }, "set": "v0.1", "judge": "Claude Opus", "n": "8 items, 16 judge calls", "seeds": "1 (control s44; arm deterministic)", "result": "recall 70 % < 75 % bar", "result_verified": true, "effect": "recall 70 % < 75 % bar", "ci": null, "mde": "n/a", "verdict": "adopted", "verdict_detail": "lever-adopted (ADOPT by frozen criterion, research-only; reached the claim build via r6 in cut 7)", "reopen_if": "recall axis only: see FACTS-RECALL.md and LCRELAX-PROBE.md", "closed_row": "R9", "registry_public": "boar-research:reports/CLOSED.md:24", "date": "2026-10-06", "fact_ids": [ "F-421" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-046", "code": "FACTS-RECALL", "name": "FACTS-RECALL", "family": "4B-fast-path", "question": "What are FACTS r2 trigger recall and precision on a blind fresh-phrasing set of 40 fact questions and 20 near-misses?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "recall-set.jsonl (FACTS blind set 1)", "judge": "none (deterministic gate)", "n": "60 (40 fact + 20 near-miss)", "seeds": "deterministic", "result": "recall 70 % < 75 % bar", "result_verified": true, "effect": "recall 70 % < 75 % bar", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed (measurement only; set became dev after r3)", "reopen_if": "recall axis only: see FACTS-RECALL.md and LCRELAX-PROBE.md", "closed_row": "R9", "registry_public": "boar-research:reports/CLOSED.md:24", "date": "2026-10-06", "fact_ids": [ "F-421" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-047", "code": "FACTS-r2", "name": "FACTS-r2", "family": "4B-fast-path", "question": "Do the DATA-REVIEW section 4 build-time data gates and stricter trigger rules (r2) remove the probed wrong-answer paths while keeping the 8 target answers and gates G1-G5 identical to r1?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/facts/FACTS-criterion.md" }, "set": "v0.1; dev chat v0-v2 (192); safety track; V-exact 11", "judge": "none (deterministic gate)", "n": "13 DATA-REVIEW probes + 39 wider probes; 8 targets; 52/12/11/192 G3 items", "seeds": "deterministic", "result": "recall 70 % < 75 % bar", "result_verified": true, "effect": "recall 70 % < 75 % bar", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed (gated rework passed; superseded by r3)", "reopen_if": "recall axis only: see FACTS-RECALL.md and LCRELAX-PROBE.md", "closed_row": "R9", "registry_public": "boar-research:reports/CLOSED.md:24", "date": "2026-10-06", "fact_ids": [ "F-421" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-048", "code": "FACTS-r4", "name": "FACTS-r4", "family": "4B-fast-path", "question": "Do a safe-lowercase multi-word name rule and a typo-tolerant 'the' raise FACTS recall to >=85% on blind set 3 with 0 false positives vs r3?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/facts/FACTS-r4-criterion.md" }, "set": "recall-set.jsonl (dev); recall-set-2.jsonl (dev); probes-r4-fresh (22)", "judge": "none (deterministic gate)", "n": "sets 1 and 2 (60 each, dev); 22 fresh lowercase-homonym probes", "seeds": "deterministic", "result": "recall 70 % < 75 % bar", "result_verified": true, "effect": "recall 70 % < 75 % bar", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed (FAIL on the fresh-precision gate before the blind set; set 3 untouched)", "reopen_if": "recall axis only: see FACTS-RECALL.md and LCRELAX-PROBE.md", "closed_row": "R9", "registry_public": "boar-research:reports/CLOSED.md:24", "date": "2026-10-06", "fact_ids": [ "F-421" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-049", "code": "FACTS-r5", "name": "FACTS-r5", "family": "4B-fast-path", "question": "Does a two-sided neighbour guard (plus openers and whole-label override) fix lowercase-homonym precision while reaching >=85% recall on blind set 3 (vs r3/r4)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/facts/FACTS-r5-criterion.md" }, "set": "recall-set-3.jsonl (FACTS blind set 3); nearmiss-lc.jsonl (blind adversarial lowercase); recall-set.jsonl (dev); recall-set-2.jsonl (dev)", "judge": "none (deterministic gate)", "n": "set 3: 60 (40 fact + 20 near-miss); nearmiss-lc 40; dev sets 1-2", "seeds": "deterministic", "result": "recall 70 % < 75 % bar", "result_verified": true, "effect": "recall 70 % < 75 % bar", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed (FAIL on blind set 3: recall 60 % < 75 %)", "reopen_if": "recall axis only: see FACTS-RECALL.md and LCRELAX-PROBE.md", "closed_row": "R9", "registry_public": "boar-research:reports/CLOSED.md:24", "date": "2026-10-06", "fact_ids": [ "F-421" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-050", "code": "FACTS-r6", "name": "FACTS-r6", "family": "4B-fast-path", "question": "Does asset coverage expansion (1,036 -> 4,233 entities) plus 5 trigger fixes on r5 reach >=85% recall on blind set 4 with 0 false positives?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/facts/FACTS-r6-criterion.md" }, "set": "recall-set-4.jsonl (FACTS blind set 4); nearmiss-lc2.jsonl (blind adversarial lowercase); recall sets 1-3 (dev); nearmiss-lc (dev)", "judge": "none (deterministic gate)", "n": "set 4: 60 (40 fact + 20 near-miss); nearmiss-lc2 40; dev sets 1-3", "seeds": "deterministic", "result": "recall 70 % < 75 % bar", "result_verified": true, "effect": "recall 70 % < 75 % bar", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed: FAIL on blind set 4 (recall 70 % < 75 %); r6 kept as head by coordinator and shipped in cut 7 via CUT7 (0 fires on v0.2/V-para)", "reopen_if": "recall axis only: see FACTS-RECALL.md and LCRELAX-PROBE.md", "closed_row": "R9", "registry_public": "boar-research:reports/CLOSED.md:24", "date": "2026-10-06", "fact_ids": [ "F-421" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-051", "code": "FOLLOWQ", "name": "FOLLOWQ", "family": "retrieval", "question": "Retrieval-only diagnostic on cut 5: does a history-aware query (Q1 previous turn + current, Q2 carried entities) add sources on >=8 rule-flagged follow-ups while changing retrieval on <=2 non-follow-ups, vs Q0 (bare query)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "dev chat slices v0-v2 (CHATGATE 192)", "judge": "none (deterministic gate)", "n": "192 dev chat messages (21 follow-ups + 171 non-follow-ups; 69 expected-search); 576 requests", "seeds": "seed 44", "result": "gains 10/21 follow-ups but changes 11/171 others", "result_verified": true, "effect": "gains 10/21 follow-ups but changes 11/171 others", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed (diagnostic decision rule NOT met)", "reopen_if": "a variant that leaves non-follow-up retrieval unchanged (<=2/171)", "closed_row": "R4", "registry_public": "boar-research:reports/CLOSED.md:19", "date": "2026-10-06", "fact_ids": [ "F-514" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-052", "code": "FOLLOWQG", "name": "FOLLOWQG", "family": "retrieval", "question": "Does cut 5 + Q1 history-aware query applied only when the FOLLOWQ follow-up rule fires produce better follow-up answers (Opus A/B, >=60% non-tie wins) vs cut 5, with zero retrieval change elsewhere?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "dev chat slices v0-v2 (192); v0.1", "judge": "Claude Opus", "n": "13 changed follow-up answers (26 A/B decisions); 192 chat + 60 v0.1 for identity", "seeds": "seed 44", "result": "45 % wins (bar 60 %)", "result_verified": true, "effect": "45 % wins (bar 60 %)", "ci": null, "mde": "n/a (A/B gate)", "verdict": "adopted", "verdict_detail": "closed (NOT ADOPTED: G3 FAIL 45 % < 60 %)", "reopen_if": "none stated (R4 family)", "closed_row": "R63", "registry_public": "boar-research:reports/CLOSED.md:92", "date": "2026-10-06", "fact_ids": [ "F-543" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-053", "code": "GATEFID", "name": "GATEFID", "family": "tooling", "question": "Does running the phone's route gate (nextTokenProbs) on desktop change cut 4 / cut 5 answers vs a gate-less cut 4 control on v0.1 EN s44?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN", "judge": "none (deterministic gate)", "n": "60", "seeds": "s44", "result": "0.000 by construction (0/60 text changes, nothing judged)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed: FIDELITY OK (tooling/harness fidelity check, no judge)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-054", "code": "HFU", "name": "HFU", "family": "4B-fast-path (health routing diagnostic)", "question": "Deterministic diagnostic on cut 5 (wt/cut5lite 7173c9af): how often does a health question asked as an elliptical/pronoun follow-up lose the health route (cards / health route / isHealthFollowUp / danger detector) that the same question gets standalone; decision rule: build a fix arm if >= 10% of follow-up forms lose the route.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "dev chat slices v0-v2 (192 conversations); emergency-dev (the evaluation/safety research agent dev.jsonl); bench v0.1 public health/danger items; water-safety dev probes", "judge": "none (deterministic gate)", "n": "Set A: 192 dev chat conversations (18 health-matching, 11 follow-up turns, baseline n = 3); Set B: 111 dev probes -> 22…", "seeds": "n/a (deterministic, no generation)", "result": "n/a (diagnostic: 69/100 synthetic follow-up forms lose the standalone health route; combined 71/103 = 69%; EN 33/58 = 57%)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "diagnostic complete (closed): decision rule (>= 10% lost) met, 69/100, so the fix arm HFU-FIX was built", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-055", "code": "HFUFIX", "name": "HFUFIX", "family": "4B-fast-path (health routing)", "question": "Cut 5 + one general rule (short pronoun/opener follow-up whose condition is recovered from the last 2 user turns is routed like the standalone question) vs cut 5 control (rule off): routing-only gates G1-G5 incl. held-out template recall >= 90% (G3) and 0 false health routes (G5).", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN s44 (identity); dev chat slices v0-v2 (192); HFU synthetic set (EN dev health probes); v0.1 EN non-health items (G5b pairs)", "judge": "none (deterministic gate)", "n": "G3: 58 EN dev probes x 5 templates, 248 forms with health standalone; G5b: 20 non-health pairs; G2: v0.1 EN 60 + 192 de…", "seeds": "v0.1 EN s44 (G2 identity); otherwise deterministic", "result": "58 % / 84 % vs bar 90 %", "result_verified": true, "effect": "58 % / 84 % vs bar 90 %", "ci": null, "mde": "n/a (gate)", "verdict": "adopted", "verdict_detail": "closed: NOT ADOPTED as frozen (G3 58.1% < 90%; G5b 4/20 vs 0, pre-existing). The rescue HFU-FIX2 also failed, so the family is closed with no further rescue (QUEUE.md:690).", "reopen_if": "see R62", "closed_row": "R61", "registry_public": "boar-research:reports/CLOSED.md:90", "date": "2026-10-06", "fact_ids": [ "F-542" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-056", "code": "HFUFIX2", "name": "HFUFIX2", "family": "4B-fast-path (health routing)", "question": "Re-test of the same frozen HFUFIX patch on fresh templates T2 with post-hoc declared scope (card-condition probes only; G5b counts only arm-added routes) vs control (rule off): G3' >= 90% same route and arm >= control + 30 points; G5b' 0 added routes; one shot, no further rescue.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "EN dev health probes with card condition (HFU set); v0.1 EN non-health items (G5b' pairs)", "judge": "none (deterministic gate)", "n": "19 of 58 EN dev probes with a card condition; 76 forms whose standalone routes health; 20 non-health pairs", "seeds": "n/a (deterministic routing)", "result": "58 % / 84 % vs bar 90 %", "result_verified": true, "effect": "58 % / 84 % vs bar 90 %", "ci": null, "mde": "n/a (gate)", "verdict": "adopted", "verdict_detail": "closed: NOT ADOPTED. The HFU-FIX family is closed with no further rescue (HFUFIX2.md:5; QUEUE.md:690).", "reopen_if": "none (criterion: no further rescue)", "closed_row": "R62", "registry_public": "boar-research:reports/CLOSED.md:91", "date": "2026-10-06", "fact_ids": [ "F-542" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-057", "code": "L1-PROTOCOL", "name": "L1-PROTOCOL (judge triage protocol)", "family": "judge-protocol", "question": "Define the L1 triage judge protocol (slice, judge, sharding, promotion/futility rules) and its addenda (HAIKUJUDGE, PARJUDGE, OPUSSHARD, SLICE20, near-miss, SEEDVAR, RANDOM20).", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1 EN; v0.2; V-para", "judge": "Claude Opus, Claude Haiku (L1), Claude Sonnet", "n": "n/a", "seeds": "n/a", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling (protocol frozen 2026-10-06; addenda HAIKUJUDGE/PARJUDGE 10-07, OPUSSHARD/SLICE20 10-08, near-miss/SEEDVAR 10-10, RANDOM20 labelled 10-11)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [], "chapter": "measurer", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-058", "code": "NAMEDSLOT", "name": "NAMEDSLOT", "family": "retrieval", "question": "Does giving a question-named low-share title one lead passage after the first hits (cut 6 + one searchDetailed rule) raise gold recall without disturbing retrieval breadth, vs cut 6 retrieval replay?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 60", "judge": "none (deterministic gate)", "n": "60 (v0.1 EN; 39 gold items)", "seeds": "replay of s44 (deterministic)", "result": "recall@5 0.333 → 0.410 but changed 15/60 lists (bar ≤ 12)", "result_verified": true, "effect": "recall@5 0.333 → 0.410 but changed 15/60 lists (bar ≤ 12)", "ci": null, "mde": "n/a (deterministic gate)", "verdict": "closed", "verdict_detail": "closed (NOT ADOPTED as frozen: G2c breadth gate FAIL 15/60 vs ≤ 12; G3/G4 not run by stop rule)", "reopen_if": "a rule that moves <=12/60 raw lists (generic words resolving to travel guides were the failure)", "closed_row": "R3", "registry_public": "boar-research:reports/CLOSED.md:18", "date": "2026-10-06", "fact_ids": [ "F-513", "F-502" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-059", "code": "SIGNAL20", "name": "SIGNAL20 (signal slices)", "family": "judge-protocol", "question": "Fix a 20-item L1 triage slice (20 lowest 4B correctness) on v0.1 EN, and later on v0.2 and V-para from the cut 6 claim run, for fast paired judge triage.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN; v0.2 (bench-v02) ids; V-para (V-para) ids", "judge": "Claude Opus", "n": "20 per set", "seeds": "control s44 (v0.1); cut 6 claim s42+s43 (v0.2 / V-para)", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling (v0.1 slice frozen 10-06; held-out-style slices added 10-08; retired for arms frozen after the 10-11 RANDOM20 addendum)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [], "chapter": "honest-number", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-060", "code": "SPECDEC", "name": "SPECDEC", "family": "device", "question": "Does speculative decoding with a small draft (Qwen3-0.6B Q8_0, or the kit's Qwen2.5-1.5B) on the 4B answer path pass identity, projected-A15-saving, RAM and llama.rn-path gates vs non-speculative decoding, on the cut 5 v0.1 EN s44 answer prompts.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1", "judge": "none (deterministic gate)", "n": "50 model-path prompts (of 60 v0.1 EN s44 items)", "seeds": "1 (s44 prompts; engine sampler seed 44 + greedy twin)", "result": "acceptance 31 %; projected saving 8.4 % (bar 20 %)", "result_verified": true, "effect": "acceptance 31 %; projected saving 8.4 % (bar 20 %)", "ci": null, "mde": "n/a (identity/projection gates)", "verdict": "closed", "verdict_detail": "closed (NO-GO; G1–G4 all FAIL for Qwen3-0.6B; Qwen2.5-1.5B refused on vocab)", "reopen_if": "a draft with acceptance ≥ ≈ 60 % at n_max 4, e.g. a 4B MTP head (SPECDEC.md:69)", "closed_row": "R66", "registry_public": "boar-research:reports/CLOSED.md:95", "date": "2026-10-06", "fact_ids": [ "F-545", "F-808" ], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-061", "code": "SPEEDPACK", "name": "SPEEDPACK", "family": "4B-fast-path", "question": "Is the faithful prompt diet (−126 fixed tokens) plus a 96-token answer cap on the 4B fast path non-inferior to the same-day cut 4 control on v0.1 EN s44+s45, while saving device latency.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1", "judge": "Claude Opus", "n": "57 judged items x 2 seeds (60 items per seed; 50 changed per seed)", "seeds": "2 (s44, s45)", "result": "quality −0.016; confident-wrong 5 → 9", "result_verified": true, "effect": "quality −0.016; confident-wrong 5 → 9", "ci": null, "mde": "~0.036", "verdict": "closed", "verdict_detail": "closed (NOT ADOPTED: 3 of 5 gates FAIL)", "reopen_if": "a new device-day latency experiment (patch 9904d4e1 kept); see QUEUE l.614, l.660", "closed_row": "R7", "registry_public": "boar-research:reports/CLOSED.md:22", "date": "2026-10-06", "fact_ids": [ "F-516", "F-1097" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-062", "code": "T1-DEEP35", "name": "T1-DEEP35", "family": "deep-tier", "question": "Draft: does cut 6 + the deep35 (Qwen3.6-35B-A3B) reader on research/compare routes beat the cut-6 4B claim runs on V-para (floor-adj Δ >= +0.05, CI90 low > 0) and v0.2 (raw CI90 low > 0.50), to justify reopening the deep tier for device work?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "V-para; v0.2", "judge": "Claude Opus", "n": "proposed: V-para 51 judged + v0.2 110", "seeds": "proposed s42 + s43", "result": "deep35 vs 4B: 0.565 vs 0.505, Δ +0.060 (+0.021…+0.101) raw", "result_verified": true, "effect": "deep35 vs 4B: 0.565 vs 0.505, Δ +0.060 (+0.021…+0.101) raw", "ci": "+0.021…+0.101", "mde": null, "verdict": "closed", "verdict_detail": "superseded draft: never authorised or run; replaced by DEEPVPARA (criterion frozen 2026-10-08)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-06", "fact_ids": [ "F-033" ], "chapter": "honest-number", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-063", "code": "CRYPTOFACTS-step1", "name": "CRYPTOFACTS-step1", "family": "4B-fast-path (FACTS-style instant tier / crypto)", "question": "Coverage analysis (0 model/judge calls): would a deterministic structured-fact instant tier for crypto (protocol constants, dates, EIP metadata; no time-varying data) answer enough of the 11 v0.1 crypto items where the 4B (control spctl-v01-s44) fails to justify a step-2 prototype?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 (11 crypto items); signal20 (7 crypto ids)", "judge": "none (deterministic gate)", "n": "11 v0.1 crypto items (7 in signal20)", "seeds": "n/a (4B correctness read from existing control spctl-v01-s44)", "result": "2/11 vs ≥ 6", "result_verified": true, "effect": "2/11 vs ≥ 6", "ci": null, "mde": "n/a (coverage gate)", "verdict": "closed", "verdict_detail": "closed (step-1 coverage gate NO; 'Line closed at step 1', QUEUE l.796)", "reopen_if": "none stated; QUEUE: crypto losses are explanation/reasoning, not lookups", "closed_row": "R51", "registry_public": "boar-research:reports/CLOSED.md:80", "date": "2026-10-07", "fact_ids": [ "F-538" ], "chapter": "agent-lab", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-064", "code": "CUT7", "name": "CUT7", "family": "claim-build", "question": "Does adding FACTS r6 to the frozen cut 6 build change any v0.2 held-out or V-para answer, and if so are changed answers better with zero regressions and zero wrong fires (control = cut 6 claim-run traces/verdicts)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02); V-para (V-para); v0.1 public + safety track; dev chat 192; V-exact 11", "judge": "none (deterministic gate)", "n": "v0.2 110, V-para 58, v0.1 public + safety 72", "seeds": "s42 queries (s43 verified identical); no LLM traces", "result": "2026-10-06 rule: only routing/safety fixes with v0.1 answers unchanged; 2026-10-07: cut 7 = cut 6 + FACTS r6 changed 8 v0.1 answers (0 fires on the held-out sets)", "result_verified": true, "effect": "2026-10-06 rule: only routing/safety fixes with v0.1 answers unchanged; 2026-10-07: cut 7 = cut 6 + FACTS r6 changed 8 v0.1 answers (0 fires on the held-out sets)", "ci": null, "mde": "n/a", "verdict": "adopted", "verdict_detail": "lever-adopted (claim build cut 7, adopted on gates alone under the zero-fire clause, CUT7-criterion.md:11)", "reopen_if": "recall axis only: see FACTS-RECALL.md and LCRELAX-PROBE.md", "closed_row": "R9", "registry_public": "boar-research:reports/CLOSED.md:24", "date": "2026-10-07", "fact_ids": [ "F-046", "F-421" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-065", "code": "DEEPPAR", "name": "DEEPPAR", "family": "device", "question": "D4 parity (a)+(b): are prompt tokens equal and greedy first-64-token outputs identical between bmoe v0.28.0 and llama-server b11010 on 10 deep35 probe prompts?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 (deep35 probe prompts, dev)", "judge": "none (deterministic gate)", "n": "10 prompts", "seeds": "greedy (n/a)", "result": "0.623 vs 0.634, Δ −0.012 (−0.045…+0.020); greedy token-identical 0/10", "result_verified": true, "effect": "0.623 vs 0.634, Δ −0.012 (−0.045…+0.020); greedy token-identical 0/10", "ci": "−0.045…+0.020", "mde": "0.055", "verdict": "parity", "verdict_detail": "closed — FAIL (D4 (a)+(b)); superseded for planning by DEEPPAR-L1 TRANSFER", "reopen_if": "none: D2/D3/D5 wait for a device", "closed_row": "R16", "registry_public": "boar-research:reports/CLOSED.md:35", "date": "2026-10-07", "fact_ids": [ "F-037", "F-813" ], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-066", "code": "DEEPPAR-L1", "name": "DEEPPAR-L1", "family": "device", "question": "D4 step (c): at the judged level, does bmoe v0.28.0 produce answers of the same quality as llama-server b11010 (control A deep35 R450; B = 4B R450) on identical R450 gold-passage prompts?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 57 (R450 oracle contexts)", "judge": "Claude Opus", "n": "57", "seeds": "1 (seed 42, temp 0.7)", "result": "0.623 vs 0.634, Δ −0.012 (−0.045…+0.020); greedy token-identical 0/10", "result_verified": true, "effect": "0.623 vs 0.634, Δ −0.012 (−0.045…+0.020); greedy token-identical 0/10", "ci": "−0.045…+0.020", "mde": "0.055", "verdict": "parity", "verdict_detail": "closed — TRANSFER (judged parity; equivalence shown, not a lever)", "reopen_if": "none: D2/D3/D5 wait for a device", "closed_row": "R16", "registry_public": "boar-research:reports/CLOSED.md:35", "date": "2026-10-07", "fact_ids": [ "F-037", "F-813" ], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-067", "code": "DEEPROUTE", "name": "DEEPROUTE", "family": "deep-tier", "question": "Descriptive (0 calls): where does the deep tier (Qwen3.6-35B-A3B engine, night-10-01) beat the 4B engine item by item on v0.1 EN 57, and what would a simulated hybrid router capture?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 57", "judge": "Claude Opus", "n": "57", "seeds": "deep s42/s43; 4B s42/s43/s44 (existing rows)", "result": "43 %/34 % capture out-of-sample", "result_verified": true, "effect": "43 %/34 % capture out-of-sample", "ci": null, "mde": "0.03", "verdict": "parity", "verdict_detail": "closed (descriptive, no verdict rule; routing family closed in CLOSED R28)", "reopen_if": "a pre-generation signal capturing >=70 % of the gain at <=40 % deep (none tested reached it)", "closed_row": "R28", "registry_public": "boar-research:reports/CLOSED.md:47", "date": "2026-10-07", "fact_ids": [ "F-814" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-068", "code": "FTDATA-1K", "name": "FTDATA-1K", "family": "data/finetune", "question": "Does the 1k grounded-reading SFT set generated with the three pilot-driven prompt changes clear filter >=0.70, disclaimers <=10%, and a strict faithfulness review <=2/20?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "local knowledge packs (enwiki 520, enwikivoyage 200, crypto 140, usgov 140)", "judge": "Claude Opus, independent reviewer agent", "n": "1,000 generated; 982 passing; 20 reviewed", "seeds": "seed 47 (sample), seed 48 (review sample)", "result": "unsupported-claim rate 16/20 → 11/20 → human 6/20; bar ≤ 2/20", "result_verified": true, "effect": "unsupported-claim rate 16/20 → 11/20 → human 6/20; bar ≤ 2/20", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed (FAIL on the strict review bar; LoRA not authorised)", "reopen_if": "a bar pre-registered relative to the human baseline (<=6/20 pooled CLAPnq+HAGRID, fresh seed, same 17+3 mix) or a quote-then-answer pilot; only with the maintainer", "closed_row": "R12", "registry_public": "boar-research:reports/CLOSED.md:27", "date": "2026-10-07", "fact_ids": [ "F-520" ], "chapter": "walls", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-069", "code": "FTDATA-PILOT", "name": "FTDATA-PILOT", "family": "data/finetune", "question": "Is the grounded-reading SFT data pipeline (Opus teacher on pack passages + deterministic fidelity filter) viable on 50 examples: filter pass >=60%, <=120 s/example, <=1/10 unsupported claims?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "local knowledge packs (enwiki, enwikivoyage, crypto, usgov preparedness)", "judge": "Claude Opus, independent reviewer agent", "n": "50 examples; 10 reviewed", "seeds": "seed 46", "result": "unsupported-claim rate 16/20 → 11/20 → human 6/20; bar ≤ 2/20", "result_verified": true, "effect": "unsupported-claim rate 16/20 → 11/20 → human 6/20; bar ≤ 2/20", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed (VIABLE by the letter; 1k proposed, then run as FTDATA-1K)", "reopen_if": "a bar pre-registered relative to the human baseline (<=6/20 pooled CLAPnq+HAGRID, fresh seed, same 17+3 mix) or a quote-then-answer pilot; only with the maintainer", "closed_row": "R12", "registry_public": "boar-research:reports/CLOSED.md:27", "date": "2026-10-07", "fact_ids": [ "F-520" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-070", "code": "FTDATA-REJ", "name": "FTDATA-REJ", "family": "data/finetune", "question": "Can a Sonnet-low sentence-level rejection pass (delete unsupported sentences, no regeneration) turn ftdata-1k into a set with >=600 kept examples and a strict review <=2/20?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "ftdata-1k.jsonl (982 passing rows)", "judge": "Claude Opus, Claude Sonnet, independent reviewer agent", "n": "982 rows; 1,002 verifier calls; 20 reviewed", "seeds": "seed 50 (review sample)", "result": "unsupported-claim rate 16/20 → 11/20 → human 6/20; bar ≤ 2/20", "result_verified": true, "effect": "unsupported-claim rate 16/20 → 11/20 → human 6/20; bar ≤ 2/20", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed (FAIL; grounded-reading fine-tune DATA line closed for now)", "reopen_if": "a bar pre-registered relative to the human baseline (<=6/20 pooled CLAPnq+HAGRID, fresh seed, same 17+3 mix) or a quote-then-answer pilot; only with the maintainer", "closed_row": "R12", "registry_public": "boar-research:reports/CLOSED.md:27", "date": "2026-10-07", "fact_ids": [ "F-520" ], "chapter": "agent-lab", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-071", "code": "HAIKUJUDGE", "name": "HAIKUJUDGE", "family": "judge-protocol", "question": "Does Claude Haiku 5.5 (effort low) agree with Opus judge.v1 at least as well as Sonnet-low on deep35-R450 and 4B-R450 so it can replace Sonnet-low as the L1 judge?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN", "judge": "Claude Opus, Claude Haiku (L1), Claude Sonnet", "n": "57 items x 2 legs (228 Haiku calls)", "seeds": "judge seed 7; bootstrap seed 5", "result": "Haiku ratio 0.566 vs Opus 0.634, Δ −0.069 (95 % CI −0.090, −0.049). Ranking deep35 − 4B: Haiku +0.007 [−0.018, +0.032] vs Opus +0.029 [−0.004, +0.063].", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (judge protocol: Haiku 5.5-low adopted as L1 judge, 4/4 bars; addendum applied to L1-PROTOCOL.md:11)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-07", "fact_ids": [ "F-1215" ], "chapter": "measurer", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-072", "code": "HUMANSFT", "name": "HUMANSFT", "family": "data/finetune", "question": "Data screen: do human-verified grounded-QA targets (CLAPnq, HAGRID, Dolly closed_qa) pass numeric fidelity >= 0.98 and the strict faithfulness review (<= 2/20 flagged) that the Opus teacher set failed (REVIEW-20: 16/20), so they can be LoRA targets?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "PrimeQA/clapnq train; miracl/hagrid train; databricks-dolly-15k closed_qa", "judge": "Claude Opus, independent reviewer agent", "n": "pool 4,601 rows (CLAPnq 1,811 ans + 1,748 unans, HAGRID 278, Dolly 764); review 20 rows (17 CLAPnq + 3 HAGRID)", "seeds": "review sample seed 51", "result": "unsupported-claim rate 16/20 → 11/20 → human 6/20; bar ≤ 2/20", "result_verified": true, "effect": "unsupported-claim rate 16/20 → 11/20 → human 6/20; bar ≤ 2/20", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed: FAIL (strict review 6/20 vs bar <= 2/20; numeric fidelity also failed for HAGRID 0.9640 and Dolly 0.8482)", "reopen_if": "a bar pre-registered relative to the human baseline (<=6/20 pooled CLAPnq+HAGRID, fresh seed, same 17+3 mix) or a quote-then-answer pilot; only with the maintainer", "closed_row": "R12", "registry_public": "boar-research:reports/CLOSED.md:27", "date": "2026-10-07", "fact_ids": [ "F-520", "F-503" ], "chapter": "walls", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-073", "code": "JUDGE2", "name": "JUDGE2", "family": "judge-protocol", "question": "Calibration: how closely does a local non-Claude, non-Qwen judge (Gemma 4 26B-A4B UD-Q2_K_XL) running judge.v1 bytes agree with Opus judge.v1 (and Sonnet-low) on v0.1 EN 57, and does it preserve the deep35 > 4B ranking? No GO bar.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 57 (DEEPREAD deep35-R450, qwen3-4b-R450)", "judge": "Claude Opus, Claude Sonnet, local Gemma judge", "n": "57 items per leg x 2 orders; 228 calls (leg A deep35-R450 114, leg B 4B-R450 114)", "seeds": "judge seed 7, temp 0; bootstrap seed 5", "result": "+0.026 (−0.001…+0.058) deep35 − 4B under Gemma vs +0.029 (−0.004…+0.063) under Opus; Δratio vs Opus +0.110 (+0.085…+0.134) leg A / +0.112 (+0.082…+0.142) leg B", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "measurement (calibration, no GO bar): usable for triage/ranking, not for the headline ratio", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-07", "fact_ids": [], "chapter": "measurer", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-074", "code": "L1CTL", "name": "L1CTL", "family": "judge-protocol", "question": "Control run: re-judge the signal20 paired L1 control slice (cut 3 = cut 6 answers, v0.1 s44) with Haiku 5.5-low to produce reusable control judgments and a sanity comparison vs the Opus judge.v1-med rows.", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1 (signal20 slice)", "judge": "Claude Haiku (L1)", "n": "20 ids x 2 orders = 40 Haiku calls", "seeds": "1 (s44 traces; judge seed 7)", "result": "n/a (descriptive: Haiku-low ratio 0.334 vs Opus-med 0.340, Δ −0.006 on the 20)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling / control run (no criterion, no verdict by design)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-07", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-075", "code": "LORA-DIAG", "name": "LORA-DIAG", "family": "data/finetune", "question": "Would a 1k-example grounded-reading QLoRA on Qwen3-4B-Instruct-2507 raise completeness on the shipped Q4 artifact (runtime LoRA) without losing faithfulness, numeric fidelity or safety, vs the 4B R450 control and a length-control prompt arm.", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1 EN 57 (DEEPREAD R450 protocol, planned)", "judge": "Claude Opus, Claude Sonnet", "n": "planned: 100 held-out grounded-reading examples; L1 signal20 ids; L2 57 items × 2 orders × 2 seeds", "seeds": "planned: training seed 42; L2 2 seeds", "result": "n/a (never run)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "not run: draft criterion never frozen or authorised (FTDATA 1k strict review FAIL 16/20, then the data line closed after FTDATA-REJ 11/20); superseded by SFT4B (H10)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-07", "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-076", "code": "MOESCREEN", "name": "MOESCREEN", "family": "deep-tier", "question": "Does Qwen3.5-35B-A3B UD-Q2_K_XL know more closed-book than the deep tier Qwen3.6-35B-A3B (deep35) and the 4B (GO needs Δ vs deep35 ≥ +0.03 AND vs 4B ≥ +0.05)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 57", "judge": "Claude Opus", "n": "57", "seeds": "1 (s42)", "result": "closed-book raw: Gemma 4 26B-A4B 0.613; Qwen3.5-35B-A3B 0.603; Qwen3.6-35B-A3B 0.596; vs 4B 0.487", "result_verified": true, "effect": "closed-book raw: Gemma 4 26B-A4B 0.613; Qwen3.5-35B-A3B 0.603; Qwen3.6-35B-A3B 0.596; vs 4B 0.487", "ci": null, "mde": "0.055", "verdict": "not-detected", "verdict_detail": "closed (Verdict: NO by the frozen GO rule; Δ vs deep35 +0.007 not ≥ +0.03)", "reopen_if": "a new open model <=35B total / <=4B active with a GGUF <=13 GB (BACKLOG B5)", "closed_row": "R17", "registry_public": "boar-research:reports/CLOSED.md:36", "date": "2026-10-07", "fact_ids": [ "F-034", "F-669" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-077", "code": "MOESCREEN-gemma4", "name": "MOESCREEN-gemma4", "family": "deep-tier", "question": "Does Gemma 4 26B-A4B UD-Q2_K_XL know more closed-book than deep35 (≥ +0.03) and the 4B (≥ +0.05), same MOESCREEN protocol?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 57", "judge": "Claude Opus", "n": "57", "seeds": "1 (s42)", "result": "closed-book raw: Gemma 4 26B-A4B 0.613; Qwen3.5-35B-A3B 0.603; Qwen3.6-35B-A3B 0.596; vs 4B 0.487", "result_verified": true, "effect": "closed-book raw: Gemma 4 26B-A4B 0.613; Qwen3.5-35B-A3B 0.603; Qwen3.6-35B-A3B 0.596; vs 4B 0.487", "ci": null, "mde": "0.055", "verdict": "not-detected", "verdict_detail": "closed (Verdict: NO by the frozen GO rule; Δ vs deep35 +0.018 not ≥ +0.03)", "reopen_if": "as R17", "closed_row": "R18", "registry_public": "boar-research:reports/CLOSED.md:37", "date": "2026-10-07", "fact_ids": [ "F-034", "F-670" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-078", "code": "PARJUDGE", "name": "PARJUDGE", "family": "judge-protocol", "question": "How much wall time running judge.mjs as N = 4 / N = 8 concurrent shards saves vs the sequential HAIKUJUDGE leg A1 run, and whether 429s, structured-output failures or verdict drift appear (usable = 429s <= 5 %, failures <= 2/114, same winner >= 52/57).", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 (deep35-R450 traces)", "judge": "Claude Haiku (L1)", "n": "57 items x 2 orders = 114 calls per level (228 Haiku calls total)", "seeds": "judge seed 7 (one run per level)", "result": "ratio Δ vs sequential A1: N=4 +0.001 [−0.011, +0.013]; N=8 −0.005 [−0.017, +0.007] (analysis.json dRatio 0.0009 [−0.0108, 0.0132] / −0.0051 [−0.0174, 0.0074]); wall 694 s → 154 s (4.5×) → 145 s (4.8×)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "adopted as judge protocol (tooling): both N = 4 and N = 8 USABLE; result recommends N = 8, but the L1-PROTOCOL addendum adopted default N = 4 for L1", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-07", "fact_ids": [], "chapter": "measurer", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-079", "code": "RETHURT", "name": "RETHURT", "family": "retrieval", "question": "Does real retrieval hurt the deep tier (deep35) relative to closed-book on v0.1 EN 57, and does the loss concentrate in coverage-miss / no-gold items (existing judgments only, deep35 engine vs deep35 closed-book; 4B and R450 pairs as context)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 EN 57", "judge": "Claude Opus", "n": "57 (v0.1 EN)", "seeds": "deep35 engine 42, 43; closed-book 42; 4B engine 42, 43, 44", "result": "−0.031; coverage miss carries 0.4 % of loss", "result_verified": true, "effect": "−0.031; coverage miss carries 0.4 % of loss", "ci": null, "mde": "0.055", "verdict": "closed", "verdict_detail": "closed — 'no retrieval harm shown' (pre-registered rule; coverage/no-gold share 0.4 % vs bar ≥ 60 %)", "reopen_if": "none stated", "closed_row": "R14", "registry_public": "boar-research:reports/CLOSED.md:33", "date": "2026-10-07", "fact_ids": [ "F-668" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-080", "code": "THREADS2", "name": "THREADS2", "family": "device", "question": "Would pinning llama.rn n_threads to 2 big cores (vs the engine's 4) speed 4B decode on a 2-big + 6-little Android SoC, measured on a device with the bench app?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": null, "judge": "none (deterministic gate)", "n": "proposed: 10 cut-7 prompts, ABAB × 3", "seeds": "n/a (greedy)", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "parked (dormant until a device is available; criterion not written or frozen)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-07", "fact_ids": [], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-081", "code": "VENUEGPS", "name": "VENUEGPS", "family": "tooling", "question": "Why do the 7 V-para venue items score P5 0/7 under config (b), and does injecting a per-item device location via the runner only (no engine change, cut 7) make the engine return >= 3 real venues, vs the 0/7 baseline?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para", "judge": "none (deterministic gate)", "n": "7 venue items (+5 non-venue gate controls)", "seeds": "42 (s43 not run, no pass)", "result": "P5 0/7 with injected location (s42) vs 0/7 baseline; cause shifts from no-location 7/7 to places:coverage-none 7/7", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed (diagnostic; verdict branch 0/7, runner fix not adopted by the rule; superseded by VENUEGPS-2)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-07", "fact_ids": [], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-082", "code": "VENUEGPS-2", "name": "VENUEGPS-2", "family": "tooling", "question": "With a device GPS fix AND the user's area places tile installed (kit-vexact + coordinate-selected tiles), cut 7, no engine change, do the 7 V-para venue items reach P5 (>= 3 real venues), vs VENUEGPS 0/7?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para", "judge": "none (deterministic gate)", "n": "7 venue items (+5 non-venue gate controls)", "seeds": "42, 43", "result": "P5 7/7 on s42 and 7/7 on s43 (vs 0/7 in VENUEGPS)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (claim wording + runner option; plumbing proof, no engine change; not a claim number)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-07", "fact_ids": [], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-083", "code": "DEEPROUTE-OOS", "name": "DEEPROUTE-OOS", "family": "deep-tier", "question": "Descriptive (0 calls): does the R2 routing signal ('no on-topic source kept') hold out of sample on v0.2 and V-para cut 6 claim traces (signal rate and relation to 4B correctness; no deep rows existed)?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.2 (bench-v02, burnt→dev); V-para (V-para); v0.1 EN 57 (reference)", "judge": "Claude Opus", "n": "110 (v0.2) / 51 (V-para) / 57 (v0.1 reference)", "seeds": "2 (s42+s43, majority)", "result": "43 %/34 % capture out-of-sample", "result_verified": true, "effect": "43 %/34 % capture out-of-sample", "ci": null, "mde": "0.03", "verdict": "parity", "verdict_detail": "closed (descriptive, no verdict rule; folded into routing family R28)", "reopen_if": "a pre-generation signal capturing >=70 % of the gain at <=40 % deep (none tested reached it)", "closed_row": "R28", "registry_public": "boar-research:reports/CLOSED.md:47", "date": "2026-10-08", "fact_ids": [ "F-814" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-084", "code": "DEEPSPEED", "name": "DEEPSPEED", "family": "device", "question": "Do bmoe v0.28.0 output-preserving speed knobs (K1 --ngram, K2 --ngram --draft 6, K3 --prefetch 2, K4 --ngram --prefetch 2) keep text identical and speed up deep-tier decode vs the K0 base flags, so they can enter the device P0 flags?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "DEEPPAR 10 rendered prompts (v0.1 items)", "judge": "none (deterministic gate)", "n": "10 prompts x 5 configs = 50 runs", "seeds": "greedy, one run per (config, item)", "result": "text identical 2/10 (broken)", "result_verified": true, "effect": "text identical 2/10 (broken)", "ci": null, "mde": "n/a", "verdict": "adopted", "verdict_detail": "closed: no knob adopted, P0 flags unchanged (K1/K2/K4 BROKEN 8/10 items differ, K3 IDENTICAL-NO-GAIN 0.94x)", "reopen_if": "upstream fix (arms/deepspeed/UPSTREAM-NOTE-ngram.md)", "closed_row": "R36", "registry_public": "boar-research:reports/CLOSED.md:55", "date": "2026-10-08", "fact_ids": [ "F-816", "F-810" ], "chapter": "phone", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-085", "code": "DEEPTHINK", "name": "DEEPTHINK", "family": "deep-tier", "question": "Does a 1024-token thinking budget on the deep tier (arm T: cut 6 + Qwen3.6-35B, thinking on) lift V-para vs the same deep tier without thinking (D, DEEPVPARA), and at what time cost?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.1 smoke 2 items", "judge": "Claude Opus, Claude Haiku (L1)", "n": "51 judged ids (58 requests) per seed", "seeds": "42, 43", "result": "thinking (1024) vs no thinking: raw +0.075 (+0.040…+0.110); floor-adjusted +0.093 (+0.050…+0.137)", "result_verified": true, "effect": "thinking (1024) vs no thinking: raw +0.075 (+0.040…+0.110); floor-adjusted +0.093 (+0.050…+0.137)", "ci": "+0.040…+0.110", "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (thinking on the deep tier adopted; budget later moved 1024 → 512 by DT512); L1 PROMOTED on the Δ clause", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-08", "fact_ids": [ "F-008", "F-007", "F-617", "F-1079" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-086", "code": "DEEPTHINK-2", "name": "DEEPTHINK-2", "family": "deep-tier", "question": "Does a 3072-token thinking budget (ctx 8192, arm T2) on the deep tier lift V-para over the 1024-token thinking arm T (DEEPTHINK)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.1 smoke 2 items", "judge": "Claude Haiku (L1)", "n": "51 judged ids (58 requests)", "seeds": "42", "result": "+0.010 (−0.014…+0.041)", "result_verified": true, "effect": "+0.010 (−0.014…+0.041)", "ci": "−0.014…+0.041", "mde": "0.06", "verdict": "not-detected", "verdict_detail": "closed: NO / NOT PROMOTED at L1 (budget 3072 adds nothing over 1024); CLOSED R22 labels it 'not detected at MDE 0.06 (51, 1 seed, Haiku)'", "reopen_if": "a thinking model that uses >~1.3k tokens (Qwen stops there, per CLAUDE.md)", "closed_row": "R22", "registry_public": "boar-research:reports/CLOSED.md:41", "date": "2026-10-08", "fact_ids": [ "F-525", "F-655" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-087", "code": "DEEPVPARA", "name": "DEEPVPARA", "family": "deep-tier", "question": "Does the deep tier (cut 6 + Qwen3.6-35B-A3B via llama-server, thinking off; arm D) lift the claim build's answers on V-para (stage 1) and v0.2 (stage 2) vs a same-backend 4B control C (promotion) and vs the claim traces K (claim scales: raw, floor-adjusted, win-rate)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02); v0.1 (smoke, 2 items)", "judge": "Claude Opus, Claude Haiku (L1)", "n": "51 judged V-para ids (58 generated) + 110 v0.2 ids", "seeds": "2 (42, 43)", "result": "raw 0.467 (0.439–0.494); floor-adjusted 0.329 (0.294–0.362); win 0/51; desktop p50 ≈ 33 s", "result_verified": true, "effect": "raw 0.467 (0.439–0.494); floor-adjusted 0.329 (0.294–0.362); win 0/51; desktop p50 ≈ 33 s", "ci": "0.439…0.494", "mde": null, "verdict": "measurement", "verdict_detail": "closed (measurement): Stage 1 V-para PROMOTED (Haiku Δ D−C +0.036) and Stage 2 v0.2 PROMOTED (+0.032). Under Opus the deep tier is below the bar by 0.171 on V-para and by 0.043 on v0.2; win-rate 0 on both", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-08", "fact_ids": [ "F-005", "F-006" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-088", "code": "LCRELAX", "name": "LCRELAX", "family": "4B-fast-path", "question": "Does relaxing the 'lowercase one-word names = countries/capitals only' rule per entity kind (V1-V3) raise recall on the burnt recall sets without new false positives, vs r6 (V0)?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "recall sets 1-4 (burnt, dev); nearmiss-lc (dev); nearmiss-lc2 (dev)", "judge": "none (deterministic gate)", "n": "4 recall sets x 60 + nearmiss-lc 40 + nearmiss-lc2 40 + 22 fresh probes", "seeds": "deterministic", "result": "recall 70 % < 75 % bar", "result_verified": true, "effect": "recall 70 % < 75 % bar", "ci": null, "mde": "n/a", "verdict": "closed", "verdict_detail": "closed (dev probe, not adopted; the author 'LCRELAX NO, FACTS stays at r6')", "reopen_if": "recall axis only: see FACTS-RECALL.md and LCRELAX-PROBE.md", "closed_row": "R9", "registry_public": "boar-research:reports/CLOSED.md:24", "date": "2026-10-08", "fact_ids": [ "F-421", "F-413" ], "chapter": "walls", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-089", "code": "OPUSSHARD", "name": "OPUSSHARD", "family": "judge-protocol", "question": "Can judge.mjs run Opus judge.v1 in N = 4 parallel shards without rate-limit failures or quality drift vs the sequential Opus rows (deep35-R450, 10 signal20 ids)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1", "judge": "Claude Opus", "n": "10 items x 2 orders = 20 calls", "seeds": "1 (judge seed 7)", "result": "Δratio -0.0135 (boot CI -0.025 to -0.002), 10 items × 2 orders, v0.1, Opus judge.v1", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lever-adopted (USABLE at N = 4; Opus L2/claim judging may run 4 shards; 2026-10-08 13:33)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-08", "fact_ids": [], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-090", "code": "DEEPFULL-F1", "name": "DEEPFULL-F1", "family": "deep-tier", "question": "Does the deep tier + thinking (budget 1024) gain from the app's 'detailed' persona alone (config only, cap 200, ctx 400) vs control T = DEEPTHINK with the succinct persona?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para) signal20-vpara slice", "judge": "Claude Haiku (L1)", "n": "20 (signal20-vpara slice)", "seeds": "1 (s42)", "result": "−0.066 (−0.134…+0.014)", "result_verified": true, "effect": "−0.066 (−0.134…+0.014)", "ci": "−0.134…+0.014", "mde": "0.08", "verdict": "closed", "verdict_detail": "closed: NO at the slice (Haiku only; stop after stage 2, no full set, no Opus)", "reopen_if": "none stated", "closed_row": "R24", "registry_public": "boar-research:reports/CLOSED.md:43", "date": "2026-10-09", "fact_ids": [ "F-527" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-091", "code": "DEEPFULL-F2", "name": "DEEPFULL-F2", "family": "deep-tier", "question": "Does the deep tier + thinking (budget 1024) gain from the 'detailed' persona plus fast-tier budgets (DEEP_CONTEXT_TOKENS 1200, DEEP_MAX_TOKENS 400) vs control T = DEEPTHINK with the succinct persona?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para) signal20-vpara slice; v0.1 (smoke, 2 items)", "judge": "Claude Haiku (L1)", "n": "20 (signal20-vpara slice); smoke 2 v0.1", "seeds": "1 (s42)", "result": "−0.013 (−0.069…+0.048)", "result_verified": true, "effect": "−0.013 (−0.069…+0.048)", "ci": "−0.069…+0.048", "mde": "0.08", "verdict": "not-detected", "verdict_detail": "closed: NO at the slice by the criterion rule (Haiku only; CI −0.069…+0.048 straddles 0 and includes +0.03, so CLOSED.md labels it 'not detected')", "reopen_if": "DEEPCOVER, tried and closed NO 10-12 (R67) → none planned", "closed_row": "R25", "registry_public": "boar-research:reports/CLOSED.md:44", "date": "2026-10-09", "fact_ids": [ "F-528" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-092", "code": "DEEPTHINK-V02", "name": "DEEPTHINK-V02", "family": "deep-tier", "question": "On v0.2 (110, burnt/dev), does deep + thinking (TV, budget 1024 after DEEPTHINK-2 NO) beat the deep tier without thinking (D, DEEPVPARA) at L1 Haiku?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02)", "judge": "Claude Haiku (L1)", "n": "110", "seeds": "42", "result": "raw 0.612 (0.581–0.644); floor-adjusted 0.514 (0.475–0.554); 0 wins", "result_verified": true, "effect": "raw 0.612 (0.581–0.644); floor-adjusted 0.514 (0.475–0.554); 0 wins", "ci": "0.581…0.644", "mde": null, "verdict": "closed", "verdict_detail": "NO at L1 by the frozen rule (Haiku Δ +0.027 < +0.03; stands); effect later sized positive on Opus by Addendum 1 (+0.045 raw), which prompted the L1-PROTOCOL near-miss rule", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-09", "fact_ids": [ "F-014" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-093", "code": "DEEPTHINK-V02-ADDENDUM1", "name": "DEEPTHINK-V02-ADDENDUM1", "family": "deep-tier", "question": "Claim-scale (Opus) measurement of deep + thinking 1024 (TV) on v0.2 vs D (deep, no thinking) and K (4B cut 6 claim run); measurement only, does not reopen the L1 NO.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02)", "judge": "Claude Opus", "n": "110", "seeds": "42, 43", "result": "raw 0.612 (0.581–0.644); floor-adjusted 0.514 (0.475–0.554); 0 wins", "result_verified": true, "effect": "raw 0.612 (0.581–0.644); floor-adjusted 0.514 (0.475–0.554); 0 wins", "ci": "0.581…0.644", "mde": null, "verdict": "measurement", "verdict_detail": "measurement-only (no verdict); bar reading NOT met: floor-adj 0.514 (0.475–0.554), CI low < 0.50, win-rate 0/110 ×2; point above, CI straddles on burnt/dev v0.2", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-09", "fact_ids": [ "F-014" ], "chapter": "honest-number", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-094", "code": "DT256", "name": "DT256", "family": "deep-tier", "question": "Is a 256-token deep thinking budget (arm Q) still at parity with 1024 (T, DEEPTHINK) at lower time, or is that where the thinking gain drops (with H = DT512 and D = DEEPVPARA as extra controls)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.1 (smoke, 2 items)", "judge": "Claude Opus, Claude Haiku (L1)", "n": "51 judged ids (58 generated) x 2 seeds; slice 20", "seeds": "42, 43", "result": "raw 0.514 (0.483–0.544); floor-adjusted 0.387 (0.349–0.426); vs 1024 raw −0.028 (−0.056…−0.003); p50 ≈ 55 s", "result_verified": true, "effect": "raw 0.514 (0.483–0.544); floor-adjusted 0.387 (0.349–0.426); vs 1024 raw −0.028 (−0.056…−0.003); p50 ≈ 55 s", "ci": "0.483…0.544", "mde": "0.03", "verdict": "parity", "verdict_detail": "closed (CHEAPER-LOSS)", "reopen_if": "speed-constrained device fallback only", "closed_row": "R21", "registry_public": "boar-research:reports/CLOSED.md:40", "date": "2026-10-09", "fact_ids": [ "F-011", "F-524", "F-658" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-095", "code": "DT512", "name": "DT512", "family": "deep-tier", "question": "Does halving the deep tier's thinking budget (DEEP_THINKING_BUDGET 1024 -> 512, arm H) keep the thinking gain vs T (DEEPTHINK, 1024) while cutting time (with D = DEEPVPARA no-thinking for the ladder)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.1 (smoke, 2 items)", "judge": "Claude Opus, Claude Haiku (L1)", "n": "51 judged ids (58 generated) x 2 seeds; slice 20", "seeds": "42, 43 (third seed 44 noted later via SEEDVAR)", "result": "512 vs 1024 on V-para: raw +0.011 (−0.010…+0.031), floor-adjusted +0.014 (−0.012…+0.039), ×0.55 time; on v0.2 −0.005 (−0.020…+0.009), ×0.62 time", "result_verified": true, "effect": "512 vs 1024 on V-para: raw +0.011 (−0.010…+0.031), floor-adjusted +0.014 (−0.012…+0.039), ×0.55 time; on v0.2 −0.005 (−0.020…+0.009), ×0.62 time", "ci": "−0.010…+0.031", "mde": "0.03 / 0.025", "verdict": "parity", "verdict_detail": "closed (CHEAPER-PARITY; 512 kept as a local engine recommendation, not in a claim cut)", "reopen_if": "none: budget curve closed, knee 512", "closed_row": "R20", "registry_public": "boar-research:reports/CLOSED.md:39", "date": "2026-10-09", "fact_ids": [ "F-010", "F-009", "F-523" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-096", "code": "DT512-V02", "name": "DT512-V02", "family": "deep-tier", "question": "Measurement only: what does the recommended deep configuration (thinking 512, arm H) score on v0.2 vs TV (DEEPTHINK-V02, 1024), D (deep no thinking) and K (4B cut 6 claim run)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02)", "judge": "Claude Opus, Claude Haiku (L1)", "n": "110", "seeds": "42 (seed 43 not triggered)", "result": "raw 0.608 (0.575–0.641); floor-adjusted 0.508 (0.467–0.550); win/tie/loss 0/1/109; desktop p50 ≈ 82 s", "result_verified": true, "effect": "raw 0.608 (0.575–0.641); floor-adjusted 0.508 (0.467–0.550); win/tie/loss 0/1/109; desktop p50 ≈ 82 s", "ci": "0.575…0.641", "mde": "0.03 / 0.025", "verdict": "parity", "verdict_detail": "closed (measurement only; no verdict rule)", "reopen_if": "none: budget curve closed, knee 512", "closed_row": "R20", "registry_public": "boar-research:reports/CLOSED.md:39", "date": "2026-10-09", "fact_ids": [ "F-013", "F-010", "F-523" ], "chapter": "honest-number", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-097", "code": "GEMMATHINK", "name": "GEMMATHINK", "family": "deep-tier", "question": "Does Gemma 4 26B-A4B UD-Q2_K_XL + 1024-token thinking as the deep tier (arm G) match or beat Qwen3.6-35B-A3B + thinking (control T, DEEPTHINK) on V-para, and is it faster (BETTER / PARITY-FASTER / NO)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.1 (smoke 2 items)", "judge": "Claude Opus, Claude Haiku (L1)", "n": "51 judged ids (58 generated) per seed; slice 20 ids at stage 1", "seeds": "42, 43", "result": "−0.028 (−0.057…+0.003)", "result_verified": true, "effect": "−0.028 (−0.057…+0.003)", "ci": "−0.057…+0.003", "mde": "0.03", "verdict": "closed", "verdict_detail": "closed (verdict NO: G is faster than T but not at parity. PARITY-FASTER failed because CI low −0.057 < −0.03, and BETTER failed)", "reopen_if": "none stated", "closed_row": "R19", "registry_public": "boar-research:reports/CLOSED.md:38", "date": "2026-10-09", "fact_ids": [ "F-671", "F-656" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-098", "code": "MTP", "name": "Speculative decoding: MTP", "family": "deep (device)", "question": "Speculative decoding: MTP", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": null }, "set": "external numbers, no BOAR run", "judge": null, "n": null, "seeds": null, "result": "5.6 vs 5.8–6.1 tok/s (slower)", "result_verified": true, "effect": "5.6 vs 5.8–6.1 tok/s (slower)", "ci": null, "mde": "n/a", "verdict": "measurement", "verdict_detail": "closed (external evidence; not our measurement)", "reopen_if": "a verify pass that does not widen the expert read set", "closed_row": "R35", "registry_public": "boar-research:reports/CLOSED.md:54", "date": "2026-10-09", "fact_ids": [ "F-815", "F-713", "F-808", "F-822" ], "chapter": "phone", "source": "closed registry row R35 + verified fact" }, { "id": "X-099", "code": "DEEPBO2", "name": "DEEPBO2", "family": "deep-tier", "question": "Can a judge-blind selector (S1 35B self-selection, S2 4B verifier, free heuristics) pick between the deep tier's two seeds (DT512 arm H, s42 vs s43) and capture the oracle best-of-2 gain, vs the single-seed baselines?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para", "judge": "Claude Opus", "n": "51 judged ids (V-para), 102 prompts per selector", "seeds": "existing s42 and s43 answers (no new generations); bootstrap seed 7", "result": "oracle +0.038; selectors at chance (49 %)", "result_verified": true, "effect": "oracle +0.038; selectors at chance (49 %)", "ci": null, "mde": "0.03", "verdict": "closed", "verdict_detail": "closed (NO for every selector)", "reopen_if": "a judge-free selector that beats chance", "closed_row": "R27", "registry_public": "boar-research:reports/CLOSED.md:46", "date": "2026-10-10", "fact_ids": [ "F-984", "F-659" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-100", "code": "DEEPCOV-FB", "name": "DEEPCOV-FB", "family": "retrieval", "question": "Step 0 (0-call retrieval replay on v0.1): does admitting the enwiki top-100k pack only as an R2 fallback (items where the deep tier keeps no on-topic source) put gold sources into the engine's kept set on ≥ 5 items without emptying any, vs the curated kit (base) and vs the pack as a peer (DEEPWIKI config)?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1", "judge": "none (deterministic gate)", "n": "60 v0.1 items (46 retrieving, 39 with gold, 13 R2)", "seeds": "n/a (deterministic replay)", "result": "gold enters on 1 item (gate 5)", "result_verified": true, "effect": "gold enters on 1 item (gate 5)", "ci": null, "mde": "n/a (gate)", "verdict": "closed", "verdict_detail": "closed: gate FAIL at step 0 (0 judge calls; deterministic retrieval replay on public v0.1)", "reopen_if": "none stated", "closed_row": "R32", "registry_public": "boar-research:reports/CLOSED.md:51", "date": "2026-10-10", "fact_ids": [ "F-1116", "F-1100" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-101", "code": "DEEPCTX", "name": "DEEPCTX", "family": "deep-tier", "question": "Does the deep tier + thinking 512 gain from reading 1200 context tokens (arm X, DEEP_CONTEXT_TOKENS 400 → 1200, succinct persona, cap 200) instead of 400 (control H = DT512 recommended configuration)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para) signal20-vpara slice; v0.1 (smoke, 2 items)", "judge": "Claude Haiku (L1)", "n": "20 (signal20-vpara slice); smoke 2 v0.1", "seeds": "1 (s42)", "result": "−0.006 (−0.036…+0.024)", "result_verified": true, "effect": "−0.006 (−0.036…+0.024)", "ci": "−0.036…+0.024", "mde": "0.08", "verdict": "closed", "verdict_detail": "closed: NO at the slice (Haiku only, not a near-miss; no full set, no Opus)", "reopen_if": "none planned (BACKLOG s3)", "closed_row": "R23", "registry_public": "boar-research:reports/CLOSED.md:42", "date": "2026-10-10", "fact_ids": [ "F-526", "F-660" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-102", "code": "DEEPITER", "name": "DEEPITER", "family": "deep-tier (retrieval)", "question": "Does a planner-driven second retrieval pass (≤ 2 follow-up queries, context cap 800) before the thinking call raise the deep thinking tier (arm I) vs control H = dt512 rows on v0.2?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02) random20-v02 slice", "judge": "Claude Opus, Claude Haiku (L1)", "n": "20 (random20-v02 slice)", "seeds": "1 (s42)", "result": "−0.019 (−0.046…+0.008), +32 % time", "result_verified": true, "effect": "−0.019 (−0.046…+0.008), +32 % time", "ci": "−0.046…+0.008", "mde": "0.08", "verdict": "not-detected", "verdict_detail": "closed — NO at the slice (slice-only verdict; full 110 and V-para not run) | relabelled per the 2026-10-16 outside critique: not detected at MDE 0.08 (n ≈ 20)", "reopen_if": "none stated", "closed_row": "R31", "registry_public": "boar-research:reports/CLOSED.md:50", "date": "2026-10-10", "fact_ids": [ "F-531", "F-663" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-103", "code": "DEEPNORAG", "name": "DEEPNORAG", "family": "retrieval (deep tier)", "question": "Do the knowledge packs still matter for the deep tier with thinking 512: arm N (DT512 config with config.kit = []) vs control H = DT512 slice rows on V-para (diagnostic)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para) slice 20; v0.1 (smoke, 2 items)", "judge": "Claude Haiku (L1)", "n": "20 (signal20-vpara slice)", "seeds": "1 (s42)", "result": "−0.045 (−0.083…−0.011)", "result_verified": true, "effect": "−0.045 (−0.083…−0.011)", "ci": "−0.083…−0.011", "mde": "0.08", "verdict": "closed", "verdict_detail": "closed — RETRIEVAL-MATTERS (diagnostic; stage 3 not needed since Δ ≤ −0.03)", "reopen_if": "n/a (diagnostic)", "closed_row": "R33", "registry_public": "boar-research:reports/CLOSED.md:52", "date": "2026-10-10", "fact_ids": [ "F-1117", "F-1098" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-104", "code": "DEEPRERANK", "name": "DEEPRERANK", "family": "retrieval", "question": "Step 0 (0 judge calls): does bge-reranker-v2-m3 over the engine's top-20 move gold sources into the deep tier's top-6 (and 4B top-4) on bench v0.1, vs the engine's baseline retrieve(q,k) on cut 7?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1 (60 items, public)", "judge": "none (deterministic gate)", "n": "60 (46 retrieving, 39 gold, 31 retrieving gold)", "seeds": "n/a (deterministic replay)", "result": "gold into top-6 on 2 items (gate 5)", "result_verified": true, "effect": "gold into top-6 on 2 items (gate 5)", "ci": null, "mde": "n/a (gate)", "verdict": "closed", "verdict_detail": "closed (FAIL at step 0: gold enters the deep top-6 on 2 items; gate ≥ 5; 0 judge calls)", "reopen_if": "v0.2 gold file so the gate can run on v0.2 (META item); no new mechanism", "closed_row": "R30", "registry_public": "boar-research:reports/CLOSED.md:49", "date": "2026-10-10", "fact_ids": [ "F-530", "F-662" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-105", "code": "DEEPROUTE-VPARA", "name": "DEEPROUTE-VPARA", "family": "deep-tier", "question": "Descriptive (0 calls): simulated hybrid router using R2/R8 signals from K (claim 4B cut 6) traces with the deep tier's real Opus rows (D, H = deep+think 512) on V-para (2 seeds), plus a v0.2 addendum (seed 42), vs all-4B K and all-deep.", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02)", "judge": "Claude Opus", "n": "51 (V-para) / 110 (v0.2 addendum)", "seeds": "2 on V-para (42, 43); 1 on v0.2 (42)", "result": "43 %/34 % capture out-of-sample", "result_verified": true, "effect": "43 %/34 % capture out-of-sample", "ci": null, "mde": "0.03", "verdict": "parity", "verdict_detail": "closed (descriptive, no verdict rule; routing family CLOSED R28)", "reopen_if": "a pre-generation signal capturing >=70 % of the gain at <=40 % deep (none tested reached it)", "closed_row": "R28", "registry_public": "boar-research:reports/CLOSED.md:47", "date": "2026-10-10", "fact_ids": [ "F-814" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-106", "code": "DEEPSAMP", "name": "DEEPSAMP", "family": "deep-tier", "question": "Does Qwen's vendor-recommended thinking-mode sampling (Q: temp 1.0/top_k 20/top_p 0.95/min_p 0/presence 1.5; P: temp 0.6/top_k 20/top_p 0.95/min_p 0) beat the engine default (0.7/40/0.95/0.05) on the recommended deep config H (DT512) on V-para?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para) slice 20; v0.1 (2-item smokes)", "judge": "Claude Opus, Claude Haiku (L1)", "n": "20 (V-para slice, signal20-vpara)", "seeds": "1 (seed 42)", "result": "−0.029 / −0.017", "result_verified": true, "effect": "−0.029 / −0.017", "ci": null, "mde": "0.08", "verdict": "closed", "verdict_detail": "closed (NO at the slice for both Q and P; CLOSED R26: Q closed, P borderline)", "reopen_if": "none stated", "closed_row": "R26", "registry_public": "boar-research:reports/CLOSED.md:45", "date": "2026-10-10", "fact_ids": [ "F-529", "F-661" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-107", "code": "DEEPWIKI", "name": "DEEPWIKI", "family": "packs", "question": "Does adding the enwiki top-100k full-article pack (2.54 GB) to the deep thinking tier's kit (arm W, dt512 tree) improve V-para answers vs the recommended dt512 configuration H?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para) slice 20; v0.1 (smoke, 2 items)", "judge": "Claude Haiku (L1)", "n": "20 (V-para slice, signal20-vpara)", "seeds": "1 (42)", "result": "−0.034 (−0.082…+0.003)", "result_verified": true, "effect": "−0.034 (−0.082…+0.003)", "ci": "−0.082…+0.003", "mde": "0.08", "verdict": "measurement", "verdict_detail": "closed: NO at the slice (corpus size closed for the deep tier on the desktop)", "reopen_if": "none: corpus size closed for the deep tier too", "closed_row": "R29", "registry_public": "boar-research:reports/CLOSED.md:48", "date": "2026-10-10", "fact_ids": [ "F-1115", "F-1099" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-108", "code": "LENGTHSHARE", "name": "LENGTHSHARE", "family": "judge-protocol", "question": "Measurement: how much does the 3-dim floor-adjusted ratio of H (deep + thinking 512, s42) and K (4B cut 6, s42) move on V-para when the same answers are re-judged against the frontier+web references condensed to <= 80 words, vs the original ~280-word references (control = original judgments).", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02, Part 1 dimshare only)", "judge": "Claude Opus", "n": "51 V-para judged ids, both orders, 2 arms (H, K)", "seeds": "1 (s42; re-judge only)", "result": "deep + think 512 floor-adjusted 0.412 → 0.514 (0.456–0.578) with ≤ 80-word references; paired +0.102 (+0.068…+0.144); ≈ 17 % of the gap is length, 83 % substance", "result_verified": true, "effect": "deep + think 512 floor-adjusted 0.412 → 0.514 (0.456–0.578) with ≤ 80-word references; paired +0.102 (+0.068…+0.144); ≈ 17 % of the gap is length, 83 % substance", "ci": "0.456…0.578", "mde": null, "verdict": "measurement", "verdict_detail": "closed (descriptive measurement done; no LEVER/NO verdict by design, criterion stage 4; CLOSED.md:11 lists it as a measurement)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-10", "fact_ids": [ "F-021" ], "chapter": "measurer", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-109", "code": "MP4B", "name": "MP4B", "family": "4B-fast-path", "question": "Does the engine's own multi-pass path (decompose → research → synthesize, runDeepResearch) on the 4B cut 7 beat the single-pass 4B control C on the random20-v02 slice?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02) random20 slice", "judge": "Claude Haiku (L1)", "n": "20 (random20-v02 slice)", "seeds": "1 (s42)", "result": "−0.007 (−0.036…+0.023), ×4 cost", "result_verified": true, "effect": "−0.007 (−0.036…+0.023), ×4 cost", "ci": "−0.036…+0.023", "mde": "0.08", "verdict": "closed", "verdict_detail": "closed (NO at the slice by the L1 futility rule: Δ ≤ 0 and h2h ≤ control; full 110 and Opus not run)", "reopen_if": "none stated: needs a mechanism other than decompose / research / synthesize", "closed_row": "R13", "registry_public": "boar-research:reports/CLOSED.md:28", "date": "2026-10-10", "fact_ids": [ "F-521", "F-415" ], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-110", "code": "SEEDVAR", "name": "SEEDVAR", "family": "judge-protocol / measurement (deep-tier seed noise)", "question": "How much do the recommended deep configuration's (DT512 arm H: cut 6 + Qwen3.6-35B + thinking 512) V-para numbers move with the sampling seed, adding seed 44 to s42/s43 (comparison: T, D, K two seeds; v0.2 existing rows)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02) existing rows", "judge": "Claude Opus", "n": "51 judged V-para ids (58 generated), 102 Opus calls; v0.2 110 ids existing rows", "seeds": "3 for H (s42/s43/s44); 2 for T, D, K", "result": "raw mean 0.550, SD 0.020 (0.533 / 0.572 / 0.544); floor-adjusted mean 0.432, SD 0.025 (0.412 / 0.460 / 0.425); 0 wins each seed", "result_verified": true, "effect": "raw mean 0.550, SD 0.020 (0.533 / 0.572 / 0.544); floor-adjusted mean 0.432, SD 0.025 (0.412 / 0.460 / 0.425); 0 wins each seed", "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "measurement only (no verdict rule)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-10", "fact_ids": [ "F-012", "F-020" ], "chapter": "measurer", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-111", "code": "THINK4B", "name": "THINK4B", "family": "4B-fast-path", "question": "Does the thinking twin of the claim 4B (Qwen3-4B-Thinking-2507 Q4_K_M) on the claim build's fast path, at FAST_THINKING_BUDGET 256 (arm Y) or 512 (arm Z), lift V-para answers vs the instruct 4B (C at L1, claim rows K at L2), and at what time cost?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para; v0.1 (smokes)", "judge": "Claude Opus, Claude Haiku (L1)", "n": "51 judged V-para ids (58 generated); slices 20", "seeds": "s42 only (s43 requests prepared, not run)", "result": "+0.026 (−0.013…+0.068), ×3.7 time", "result_verified": true, "effect": "+0.026 (−0.013…+0.068), ×3.7 time", "ci": "−0.013…+0.068", "mde": "0.06", "verdict": "not-detected", "verdict_detail": "closed: NO (near-lever), stopped after seed 42 (one seed, not the two-seed verdict the criterion specifies); CLOSED R10 relabels it 'not detected at MDE 0.06'", "reopen_if": "BACKLOG s3: V-para s43 (Opus 102), only if the 4B side reopens", "closed_row": "R10", "registry_public": "boar-research:reports/CLOSED.md:25", "date": "2026-10-10", "fact_ids": [ "F-518", "F-184", "F-1214" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-112", "code": "THINK4B-V02", "name": "THINK4B-V02", "family": "4B-fast-path", "question": "Is the thinking 4B at budget 512 (arm ZV, noplaces) a lever on the seed-stable v0.2 set vs the instruct 4B C (Haiku) / claim rows K (Opus)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02)", "judge": "Claude Haiku (L1)", "n": "110 (slice 20)", "seeds": "s42", "result": "+0.009 (−0.015…+0.035)", "result_verified": true, "effect": "+0.009 (−0.015…+0.035)", "ci": "−0.015…+0.035", "mde": "0.025 (Haiku compresses)", "verdict": "not-detected", "verdict_detail": "closed: NO (by rule at Haiku L1; not a near-miss); CLOSED R11 relabels it 'not detected at MDE 0.025 (110, 1 seed)'", "reopen_if": "Opus N=4 on the same 110 ids (220 calls)", "closed_row": "R11", "registry_public": "boar-research:reports/CLOSED.md:26", "date": "2026-10-10", "fact_ids": [ "F-519", "F-184" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-113", "code": "JUDGEVAR", "name": "JUDGEVAR", "family": "judge-protocol", "question": "Measure Opus judge.v1 test-retest noise (sigma_judge) by re-judging the frozen DT512 arm H s42 V-para traces (51 ids, both orders, N=4 shards) and comparing to the frozen 2026-10-09 judgments of the same traces.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para)", "judge": "Claude Opus", "n": "51 ids x 2 orders = 102 Opus calls", "seeds": "1 generation seed (s42, existing traces); judge seed 7", "result": "σ_judge ≈ 0.003 raw / 0.004 floor-adjusted; per-item r 0.96; same winner 51/51", "result_verified": true, "effect": "σ_judge ≈ 0.003 raw / 0.004 floor-adjusted; per-item r 0.96; same winner 51/51", "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "measurement complete (no verdict by design); sigma_judge ≈ 0.003 raw / 0.004 floor-adj", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-11", "fact_ids": [ "F-019", "F-1182" ], "chapter": "measurer", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-114", "code": "LOSSMAP", "name": "LOSSMAP", "family": "judge-protocol", "question": "Measurement: classify Opus judge.v1 rationales (s42, both orders) of H (deep + thinking 512) and K (4B cut 6) on v0.2 (110) and V-para (51) into loss buckets via jev, loss-weighted, to re-rank BACKLOG B2–B7.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "51 V-para + 110 v0.2 items per arm, 2 arms (322 jev requests)", "seeds": "1 (s42)", "result": "n/a (measurement)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "closed (measurement done, backlog re-ranked 10-11; per-path addendum 10-12)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-11", "fact_ids": [ "F-630" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-115", "code": "NOTES35", "name": "NOTES35", "family": "packs", "question": "Does a knowledge pack of 35B-written encyclopedic notes (Vital Articles Level 3) added to the 4B cut 7 kit reach the 4B's context (gate A) and raise the claim build, vs the curated kit only?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02) 110 ids, retrieval-only replay", "judge": "none (deterministic gate)", "n": "110 (v0.2 ids, retrieval replay); 300 notes", "seeds": "n/a (replay)", "result": "note kept 2.7 % (bar 5 %)", "result_verified": true, "effect": "note kept 2.7 % (bar 5 %)", "ci": null, "mde": "n/a (gate)", "verdict": "closed", "verdict_detail": "closed (NO at step 0: gate A failed, 3/110 kept 2.7 % < 5 % and 10/110 raw 9.1 % < 15 %; 0 judge calls)", "reopen_if": "question-shaped notes drawn from public item-independent question sets (BACKLOG B6 QNOTES)", "closed_row": "R6", "registry_public": "boar-research:reports/CLOSED.md:21", "date": "2026-10-11", "fact_ids": [ "F-515", "F-1101" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-116", "code": "RANDOM20", "name": "RANDOM20 (random stratified slices)", "family": "judge-protocol", "question": "Replace the hardest-item signal20 slices with seed-7 category-stratified random 20-id slices of v0.2 and V-para, used only as an L1 futility screen.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02) ids; V-para (V-para) ids", "judge": "none (deterministic gate)", "n": "20 per set (from 110 v0.2 and 51 V-para judged ids)", "seeds": "random.Random(7)", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling (L1 futility-screen slices; replace signal20 for arms frozen after the addendum)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-11", "fact_ids": [], "chapter": "ruler", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-117", "code": "VITALPACK-DEEP", "name": "VITALPACK-DEEP", "family": "packs", "question": "Does adding the wiki-vital5 pack (49,832 Vital Articles level-5 intros) as a plain peer pack to the deep thinking tier's kit fill empty-kept items without crowding the curated packs, vs control H (dt512 base kit)? Step-0 0-call replay gate.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1; V-para (V-para); v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "v0.1 60 (39 gold), V-para 51, v0.2 110 (replay only)", "seeds": "n/a (deterministic replay)", "result": "pack share 0.61 (bar ≤ 0.40)", "result_verified": true, "effect": "pack share 0.61 (bar ≤ 0.40)", "ci": null, "mde": "n/a (gate)", "verdict": "closed", "verdict_detail": "closed (FAIL at step 0, gate (b) crowding on v0.2 0.612 > 0.40; 0 judge calls)", "reopen_if": "reopened as VITALPACK-ADD → closed R68", "closed_row": "R34", "registry_public": "boar-research:reports/CLOSED.md:53", "date": "2026-10-11", "fact_ids": [ "F-1118", "F-1102" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-118", "code": "DEEPCOVER", "name": "DEEPCOVER", "family": "deep-tier", "question": "Does appending a key-point enumeration instruction (list 5-8 key facts in the reasoning, then answer in 120-200 words, cap 300) to the deep thinking tier (arm C = dt512 + deepcover patch) raise completeness without losing correctness vs control H (dt512: cut 6 + Qwen3.6-35B-A3B + thinking 512, succinct persona, ctx 400, cap 200)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 (smoke, 3 items); v0.2 (bench-v02) random20-v02 slice; v0.2 full 110 (generated, unjudged); V-para s42 58 (generated, unjudged)", "judge": "Claude Opus, Claude Haiku (L1)", "n": "20 (random20-v02 slice) judged; smoke 3 v0.1; full v0.2 110 and V-para s42 58 generated but unjudged", "seeds": "1 (s42)", "result": "−0.027 (−0.077…+0.019); correctness −0.11", "result_verified": true, "effect": "−0.027 (−0.077…+0.019); correctness −0.11", "ci": "−0.077…+0.019", "mde": "0.08 (slice)", "verdict": "closed", "verdict_detail": "closed: NO at the slice (Haiku +0.005 sent it to Opus under l1_futility; Opus Δ −0.027, correctness −0.11). Full v0.2 110 and V-para s42 58 generated but unjudged; verdict strength slice only (n = 20)", "reopen_if": "none: prompt-side coverage on the deep tier is closed (DEEPCOVER.md:12)", "closed_row": "R67", "registry_public": "boar-research:reports/CLOSED.md:96", "date": "2026-10-12", "fact_ids": [ "F-546", "F-664" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-119", "code": "DEEPHEDGE", "name": "DEEPHEDGE", "family": "deep-tier (prompt)", "question": "Does a 'lead with the substantive answer, safety note last' instruction appended after the persona on the deep thinking generation (arm E = dt512 + settings.deepHedge) recover the 11 v0.2 hedged-primary items, vs control H = dt512 rows?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "29 ids generated (hedged-11 ∪ random20-v02); 0 judged", "seeds": "1 (s42)", "result": "hedged 11 → 11 (items were canned health fallbacks)", "result_verified": true, "effect": "hedged 11 → 11 (items were canned health fallbacks)", "ci": null, "mde": "n/a (step-0)", "verdict": "closed", "verdict_detail": "closed at step 0 (jev gate FAIL: hedged 11 → 11; instruction fired on 0/11 hedged items)", "reopen_if": "an engine change to the health-strict no-source fallback (H7; reviewer-gated; v0.2 only)", "closed_row": "R40", "registry_public": "boar-research:reports/CLOSED.md:64", "date": "2026-10-12", "fact_ids": [ "F-533", "F-666" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-120", "code": "DENSECHUNK", "name": "DENSECHUNK", "family": "retrieval (packs)", "question": "Would chunk-level dense vectors in the format-2 packs reach the F0 fact chunk (top-12 at cos >= 0.55) on the 32 in-pack v0.2 items and put the F0 article in raw6 without gold loss, and at what pack-size/scan cost?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.2", "judge": "none (deterministic gate)", "n": "32 in-pack v0.2 (110 v0.2 simulated, 20 control)", "seeds": "1 (H s42 queries)", "result": "+0.017 (−0.000…+0.035)", "result_verified": true, "effect": "+0.017 (−0.000…+0.035)", "ci": "−0.000…+0.035", "mde": "0.03 at n 180", "verdict": "not-detected", "verdict_detail": "parked (D0 headroom PASS on part 1, FAIL on part 2 for the fused variant; superseded by DENSECHUNK-ST2, CLOSED R46, also parked)", "reopen_if": "≥ 500 pack-matched items (n for MDE 0.017), or a reader lever that converts kept chunks (GROUND-ST/STACK may carry it as a declared component)", "closed_row": "R46", "registry_public": "boar-research:reports/CLOSED.md:70", "date": "2026-10-12", "fact_ids": [ "F-1122" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-121", "code": "EVALFIX", "name": "EVALFIX", "family": "judge-protocol", "question": "T6 eval-hygiene check (no control arm): does a wins+ties rescore or alternative CI methods (logit-t, paired t) change any recorded verdict, and what v0.3 power is needed?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "1052 mapped rows; rows of 51 / 110 items", "seeds": "re-analysis of existing seeds (V-para s42/s43, v0.2 s42, SEEDVAR 3 seeds)", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling (analysis-only, 0 judge calls; no verdict changes)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-12", "fact_ids": [], "chapter": "handoff", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-122", "code": "HEALTHQ", "name": "HEALTHQ", "family": "retrieval (health path)", "question": "Does the engine's health-strict search query drop the named place/guide on v0.2 health-strict items (gate >= 3/18), explaining the danger-item pack misses?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.2", "judge": "none (deterministic gate)", "n": "18 v0.2 health-strict items", "seeds": "n/a", "result": "0/18", "result_verified": true, "effect": "0/18", "ci": null, "mde": "n/a (step 0)", "verdict": "closed", "verdict_detail": "closed (step-0 check failed its gate: 0/18 dropped vs ≥ 3; 2026-10-12 21:06)", "reopen_if": "none: misses are low-overlap questions, not a query bug; H7 HEALTHFALLBACK is a separate product item", "closed_row": "R65", "registry_public": "boar-research:reports/CLOSED.md:94", "date": "2026-10-12", "fact_ids": [ "F-544", "F-965" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-123", "code": "K2HALF", "name": "K2HALF", "family": "deep-tier", "question": "Step-0 greedy-agreement gate: does half-k expert routing (top-4 of 8 experts with decoupled top-8 normaliser, llama.cpp patch) on Qwen3.6-35B-A3B Q2_K_XL keep >=0.80 mean token agreement with stock llama.cpp on 10 public v0.1 prompts?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1 (10 DEEPPAR public ids)", "judge": "none (deterministic gate)", "n": "10 public v0.1 prompts, 64 greedy tokens, 0 judge calls", "seeds": "1 (seed 42, temp 0)", "result": "0.18 / 0.29 (bar 0.80); decode ×1.83", "result_verified": true, "effect": "0.18 / 0.29 (bar 0.80); decode ×1.83", "ci": null, "mde": "n/a (step-0)", "verdict": "not-detected", "verdict_detail": "closed: FAIL at step 0 (greedy-agreement gate); parked as a device-speed candidate", "reopen_if": "a frozen V-para slice non-inferiority test if a device plan needs ×1.8 decode", "closed_row": "R42", "registry_public": "boar-research:reports/CLOSED.md:66", "date": "2026-10-12", "fact_ids": [ "F-534", "F-963" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-124", "code": "LOCALGATE", "name": "LOCALGATE", "family": "tooling", "question": "Can a local model (Qwen3-4B-Instruct-2507 Q4_K_M; then Laya zero-shot) replace jev in the bulk boolean gates (has_fact, quote_supports), scored against cached jev verdicts on 500 sampled requests per type (rule: agreement >= 0.90 and FPR <= 0.05).", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "cached jev gate requests (wikifull A0 shard 07, wikisection/anywhere, packcond)", "judge": "none (deterministic gate)", "n": "500 requests per type (has_fact: 149 pos + 120 band + 231 neg; quote_supports: 89 pos + 120 band + 291 neg)", "seeds": "sample seed 20261012; deterministic next-token probabilities", "result": "n/a (agreement vs cached jev: Qwen has_fact 0.790 / FPR 0.103, quote_supports 0.850 / FPR 0.112; bar ≥ 0.90 / ≤ 0.05)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed (FAIL on both gate types for Qwen3-4B and for Laya zero-shot; jev stays; Laya fine-tune parked)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-12", "fact_ids": [ "F-505", "F-511" ], "chapter": "walls", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-125", "code": "NUMCHECK", "name": "NUMCHECK", "family": "claim-build (FACTS / answer-side check, deep tier)", "question": "B6: can an answer-side numeric/date check against the FACTS r6 Wikidata table cover (>= 4/16) the deep tier's (arm H s42, dt512 + think) wrong-numeric items without rewriting any correct answer (0 false rewrites)?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "V-para; v0.2", "judge": "none (deterministic gate)", "n": "16 V-para + 17 v0.2 wrong items; top-20 non-wrong per set for the false-rewrite clause", "seeds": "1 (arm H s42)", "result": "table covers 0/16, 0/17", "result_verified": true, "effect": "table covers 0/16, 0/17", "ci": null, "mde": "n/a (step-0)", "verdict": "closed", "verdict_detail": "closed (FAIL at the 0-call gate, 2026-10-12 14:21)", "reopen_if": "a table family that holds national statistics, protocol parameters, doses, historical dates and constants (≈ 5 tables for ≤ 2 items each; none fixes arithmetic)", "closed_row": "R41", "registry_public": "boar-research:reports/CLOSED.md:65", "date": "2026-10-12", "fact_ids": [ "F-422", "F-962" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-126", "code": "P4 packGround", "name": "P4 packGround (PACKCOND P4 step 0)", "family": "packs (keep/grounding, deep tier)", "question": "Does a pack-aware grounding patch (packGround: full-text onTopic, restore one compression cut, top-2 fallback) keep the F0 article on >= 16/32 v0.2 in-pack H items with <= 4/20 control kept lists changed, vs dt512 OFF?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.2; v0.1", "judge": "none (deterministic gate)", "n": "32 in-pack H v0.2 + 20 control + 39 v0.1 gold", "seeds": "n/a (replay, 0 LLM)", "result": "14/32 (bar 16)", "result_verified": true, "effect": "14/32 (bar 16)", "ci": null, "mde": "n/a (step 0)", "verdict": "closed", "verdict_detail": "closed (FAIL at the step-0 replay gate, 2026-10-12 20:37; packGround off, no judge calls; later re-tested on the 180 self-tests as GROUND-ST, CLOSED R47)", "reopen_if": "retrieval-side lever with a different gate (P4-GATE.md:25); later tried as R47 GROUND-ST and R48 PACKROUTE", "closed_row": "R69", "registry_public": "boar-research:reports/CLOSED.md:98", "date": "2026-10-12", "fact_ids": [ "F-547" ], "chapter": "measurer", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-127", "code": "PACKCOND-P1", "name": "PACKCOND-P1", "family": "packs", "question": "With the matching pack installed (fact verified in crypto/wikivoyage pack via jev), what conditional quality do H (deep+think 512) and K (4B cut 6) reach, and where along retrieve/ground/read is the fact lost?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.2; V-para", "judge": "Claude Opus, calibrated classifier (jev)", "n": "v0.2 110 (H in-pack 32, K 20); V-para 51 (6 each)", "seeds": "1 (s42)", "result": "n/a (descriptive; H v0.2 in-pack 32 above_min 0.609 (0.532–0.679) vs not-in-pack 78 0.467 (0.417–0.519), not a paired contrast)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "measurement (done; Addendum 1: all PACKCOND numbers on above_min from 10-12 21:0x)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-12", "fact_ids": [], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-128", "code": "PACKCOND-P2", "name": "PACKCOND-P2", "family": "packs", "question": "Would packs not in the kit (vital5, preparedness, shard07, enwiki top-100k) cover more V-para/v0.2 facts, and would retrieval reach and keep the covered article (H s42)?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "V-para; v0.2", "judge": "none (deterministic gate)", "n": "V-para 51, v0.2 110", "seeds": "1 (H s42)", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "measurement (done)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-12", "fact_ids": [], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-129", "code": "PACKDENSE", "name": "PACKDENSE", "family": "retrieval (packs, deep tier)", "question": "PACKCOND P4b: does a dense article router over the packs' lead vectors bring the F0 article into the raw top-6 (>= 22/32 in-pack H v0.2 items) and then lift the deep tier's in-pack conditional score vs C = H (dt512)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2; v0.1", "judge": "none (deterministic gate)", "n": "32 in-pack H v0.2 (+110 v0.2, 60 v0.1, R20 control)", "seeds": "n/a (step 0 replay/simulation)", "result": "21/32 vs gate 22", "result_verified": true, "effect": "21/32 vs gate 22", "ci": null, "mde": "n/a (step-0)", "verdict": "closed", "verdict_detail": "closed (FAIL at step 0 by projection, 2026-10-12 21:03)", "reopen_if": "an LLM query writer measured with the same gate (low prior: query2doc gains scale with model size), or pack-matched v0.3 items where the fact article is reachable by construction (PM7)", "closed_row": "R44", "registry_public": "boar-research:reports/CLOSED.md:68", "date": "2026-10-12", "fact_ids": [ "F-1120", "F-1105" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-130", "code": "QUANTCEIL", "name": "QUANTCEIL", "family": "deep-tier", "question": "Does swapping the deep thinking tier's model from Qwen3.6-35B-A3B UD-Q2_K_XL (12.29 GB) to UD-IQ3_XXS (13.2 GB), with the dt512 configuration H otherwise unchanged, move quality on v0.2 (direction) and V-para (confirmation) vs control H?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.2 (bench-v02); V-para (V-para); random20-v02 slice; v0.1 (smoke)", "judge": "Claude Opus, Claude Haiku (L1)", "n": "V-para 51 x 2 seeds; v0.2 110 (s42); slice 20; smoke 3", "seeds": "v0.2: 42; V-para: 42, 43", "result": "+0.004 / +0.008", "result_verified": true, "effect": "+0.004 / +0.008", "ci": null, "mde": "≈ 0.03 (V-para 2 seeds)", "verdict": "not-detected", "verdict_detail": "closed — NO (not detected / 'not decided at MDE 0.03' per criterion); mechanism read-out: wrong_fact axis gain on v0.2 only", "reopen_if": "a quant ≥ Q3_K_XL (16.8 GB) on hardware that holds it; nothing on the mini or a 12 GB phone", "closed_row": "R37", "registry_public": "boar-research:reports/CLOSED.md:61", "date": "2026-10-12", "fact_ids": [ "F-255", "F-208", "F-665", "F-894" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-131", "code": "RATIONALEPASS", "name": "RATIONALEPASS", "family": "judge-protocol", "question": "Over arm H's Opus judge.v1 rationales (cut 6 + Qwen3.6-35B + thinking 512, s42), which jev-classified axes (stale, numeric, date, entity swap, over-generalised, fabricated, recent, fact type) carry the loss, against pre-set DEEP-RESEARCH §3 thresholds (O-4, O-5, O-8, T3); measurement, no control arm.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "V-para 51; v0.2 110 (wrong bucket 16 and 17 items)", "seeds": "42 (single seed)", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "done — measurement (O-5 NUMCHECK met; O-4, O-8, T3 not met; entity+fabricated borderline V-para only)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-12", "fact_ids": [ "F-470" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-132", "code": "SECTIONHIT", "name": "SECTIONHIT", "family": "packs (passage selection, deep tier)", "question": "Was the fact's section/passage in the deep tier's kept context on the 32 in-pack v0.2 items, i.e. is there passage-selection headroom (keep 3 sections / alternate ranker) within kept articles?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.2", "judge": "none (deterministic gate)", "n": "32 in-pack H v0.2 (12 reading-bucket items)", "seeds": "1 (H s42 traces replay)", "result": "3/12 (gate 6/12)", "result_verified": true, "effect": "3/12 (gate 6/12)", "ci": null, "mde": "n/a (step 0)", "verdict": "closed", "verdict_detail": "closed (diagnostic gate not met → no PASSAGE arm, 2026-10-12 21:15)", "reopen_if": "none: a selection lever reaches ≤ 2/12, only with nSections ≥ 5, which conflicts with the 400-token budget (SECTIONHIT.md:26)", "closed_row": "R70", "registry_public": "boar-research:reports/CLOSED.md:99", "date": "2026-10-12", "fact_ids": [ "F-548", "F-966" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-133", "code": "SEEDSPEC", "name": "SEEDSPEC (T7)", "family": "deep-tier", "question": "Are the deep tier's (arm H, deep + think 512) wrong claims on V-para seed-specific (>=50 % of wrong items, T7 gate) so that Integrative Decoding k = 2-4 (O-6) would be worth an arm, across seeds s42/s43/s44?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "V-para (V-para)", "judge": "Claude Opus, calibrated classifier (jev)", "n": "set A 26 wrong_fact items; set B 51 judged items; 51 jev requests", "seeds": "3 (s42/s43/s44)", "result": "seed-specific wrong 4 % (bar 50 %)", "result_verified": true, "effect": "seed-specific wrong 4 % (bar 50 %)", "ci": null, "mde": "n/a (step-0)", "verdict": "closed", "verdict_detail": "closed (T7 FAILS at step 0; O-6 Integrative Decoding closed, no T9 ID arm)", "reopen_if": "a measured per-claim integration that recovers the 54–58 % 'some seed is right' cases at ≤ ×2 wall (none exists for a fused native loop)", "closed_row": "R43", "registry_public": "boar-research:reports/CLOSED.md:67", "date": "2026-10-12", "fact_ids": [ "F-535", "F-638" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-134", "code": "SYNTHDEEP", "name": "SYNTHDEEP", "family": "retrieval", "question": "Does pinning one retrieved source per side of a comparison (cmpsplit v2.1 ported onto dt512, deep thinking tier) put both sides into the kept sources on >= 5 more of the two-sided v0.1 comparison items, vs dt512 with the setting off (0-call replay gate, LOOP-PROGRAM §3 B7)?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1", "judge": "none (deterministic gate)", "n": "60 v0.1 items replayed; 7 two-sided items where the split fired", "seeds": "n/a (deterministic replay, 0 LLM calls)", "result": "+1/+3 vs bar +5", "result_verified": true, "effect": "+1/+3 vs bar +5", "ci": null, "mde": "n/a (step-0 gate)", "verdict": "closed", "verdict_detail": "closed: FAIL at step 0 (0-call replay gate; no judge call)", "reopen_if": "a kit that holds the sides' specific facts (section-level text)", "closed_row": "R38", "registry_public": "boar-research:reports/CLOSED.md:62", "date": "2026-10-12", "fact_ids": [ "F-532", "F-961" ], "chapter": "deep-tier", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-135", "code": "T-K2", "name": "T-K2", "family": "tooling", "question": "Add 4 new pack source codes (stackexchange, medlineplus, factbook, simplewiki) and optional per-row attribution to builder and engine, byte-identical for existing packs.", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1", "judge": "none (deterministic gate)", "n": "5 v0.1 queries × 2 packs → 41 hits (differential)", "seeds": "n/a", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling (done 2026-10-12; patch uncommitted in wt/packs)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-12", "fact_ids": [], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-136", "code": "T-K4a", "name": "T-K4a", "family": "packs (retrieval)", "question": "Crowding probe: added as a peer to the claim kit, does each wave-1 pack (and the 5 SE packs together) take ≤ 0.40 of kept slots without lowering kit gold on v0.2 and v0.1?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1; v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "v0.2 110 (107 retrieve) + v0.1 60 (54 retrieve)", "seeds": "n/a (0-LLM replay)", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed (probe verdict: default-eligible qa-health-se, qa-outdoors-se, boar-firstaid-en, boar-medlineplus-en; user-selected only qa-travel-se, qa-ethereum-se, qa-bitcoin-se, 5 SE together)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-12", "fact_ids": [], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-137", "code": "T-K4c", "name": "T-K4c", "family": "tooling / packs", "question": "Make pack self-tests scorable by the standard pipeline and score qa-travel-se with the 4B cut 7 + claim kit + pack (no no-pack control).", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "qa-travel-se self-test", "judge": "Claude Haiku (L1)", "n": "20", "seeds": "1 (42)", "result": "n/a (level only: raw 0.423 (0.360–0.484); later control 0.394 in SELFTEST-SCORES gives Δ +0.029 (−0.022…+0.077))", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling (pipeline + first self-test number; no control at the time)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-12", "fact_ids": [], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-138", "code": "T-K4d", "name": "T-K4d", "family": "packs (coverage)", "question": "Coverage probe: how many of the 34 V-para absent-knowledge items have their fact in the 7 wave-1 packs (jev fact-in-text + quote_supports), alone and with the kit?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "V-para (V-para)", "judge": "none (deterministic gate)", "n": "34 V-para absent items (arm H s42)", "seeds": "n/a", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed (probe: wave-1 packs cover 1/34 V-para absent items; product value, not a bar lever)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-12", "fact_ids": [], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-139", "code": "WIKIFULL-STAGE-A", "name": "WIKIFULL-STAGE-A", "family": "retrieval", "question": "Over the 34 V-para absent-knowledge items (H s42), is the missing fact present anywhere in the 15 FineWiki Aug-2025 English Wikipedia shards (jev has_fact ≥ 0.7 + quote_supports), pooled; 0-call gate.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02, dev)", "judge": "none (deterministic gate)", "n": "34 V-para (78 v0.2 dev context)", "seeds": "n/a (jev deterministic, cached)", "result": "9/34 verified", "result_verified": true, "effect": "9/34 verified", "ci": null, "mde": "n/a (0-call gate)", "verdict": "closed", "verdict_detail": "stage complete: MARGINAL 7/34 verified over 15/15 shards (intermediate; decision pooled with stage B → 9/34 MARGINAL, CLOSED R45)", "reopen_if": "a question-shaped or domain corpus that covers the comparison/explanation/synthesis misses (T-K4d: the wave-1 packs cover 1/34); or pack-matched items where coverage is by construction", "closed_row": "R45", "registry_public": "boar-research:reports/CLOSED.md:69", "date": "2026-10-12", "fact_ids": [ "F-1121" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-140", "code": "WIKISECTION", "name": "WIKISECTION", "family": "retrieval", "question": "H5: on the deep tier's on-topic absent-knowledge items, is the missing fact in a body section of the Wikipedia article it already kept (enwiki top-100k pack), bar ≥ 40 % at jev p ≥ 0.7; 0-call gate.", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "V-para 13 on-topic + 1 off-topic; v0.2 30 on-topic + 14 off-topic", "seeds": "n/a (jev; rationales from Opus s42)", "result": "fact in body 2/13, 4/30 (bar 40 %)", "result_verified": true, "effect": "fact in body 2/13, 4/30 (bar 40 %)", "ci": null, "mde": "n/a (step-0 gate)", "verdict": "closed", "verdict_detail": "closed (FAIL at 0-call gate, 2/13 V-para and 4/30 v0.2 vs bar 40 %; scope narrowed by Correction 2026-10-12; reopened in scope by WIKIFULL R45)", "reopen_if": "superseded by R45 (all Wikipedia, all 34 absent items)", "closed_row": "R39", "registry_public": "boar-research:reports/CLOSED.md:63", "date": "2026-10-12", "fact_ids": [ "F-1119", "F-1103" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-141", "code": "WIKISECTION-H5b", "name": "WIKISECTION-H5b", "family": "retrieval", "question": "H5b addendum: is the missing fact anywhere in the enwiki top-100k pack (BM25 FTS over rationale terms + kept title, jev has_fact), for absent items with a kept source.", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "V-para 13 + 1; v0.2 30 + 14", "seeds": "n/a", "result": "n/a (0-call gate; V-para 1/13 (8 %) at p ≥ 0.7, 4/13 at p ≥ 0.6; v0.2 13/30 (43 %) unverified)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed (gate FAIL on V-para 1/13 at p ≥ 0.7; reading scope-corrected 2026-10-12; superseded by WIKIFULL pooled 9/34, R45)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-12", "fact_ids": [], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-142", "code": "CROWDQ", "name": "CROWDQ", "family": "packs", "question": "What does installing all 9 published packs as peers (P, cut 7 + kit + 9 packs) cost in answer quality vs installing only the one relevant pack (S, SELFTEST-SCORES rows) on the 180 public self-test items?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "180 self-tests", "judge": "Claude Haiku (L1)", "n": "180 (9 packs × 20)", "seeds": "1 (seed 42, deterministic)", "result": "Opus −0.019 (−0.035…−0.003); Haiku −0.016 (−0.033…−0.001)", "result_verified": true, "effect": "Opus −0.019 (−0.035…−0.003); Haiku −0.016 (−0.033…−0.001)", "ci": "−0.035…−0.003", "mde": null, "verdict": "measurement", "verdict_detail": "closed (measurement): 'cost small' per the frozen strict reading (|Δ| 0.016 < 0.03, CI high < 0, not a near-miss); top-1 routing closed at the product level (confirms R48); Opus L2 later confirmed −0.019 (−0.035…−0.003)…", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-13", "fact_ids": [ "F-026" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-143", "code": "DENSECHUNK-ST", "name": "DENSECHUNK-ST", "family": "retrieval", "question": "Does a two-stage chunk-level semantic path (packDense2 fusion: lead top-50 → chunk cosine top-12 at ≥0.55, fused 0.5/0.5) in format-2 packs bring the fact passage into context and improve the shipped 4B on the 180 public pack self-test items, vs the same build with the setting off?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "180 self-tests; v0.1 public 60; v0.2 (bench-v02) 110 no-harm replay", "judge": "none (deterministic gate)", "n": "180 self-test items (+ v0.1 public 60, v0.2 110 no-harm replays)", "seeds": "n/a (replay only, 0 judge calls)", "result": "+0.017 (−0.000…+0.035)", "result_verified": true, "effect": "+0.017 (−0.000…+0.035)", "ci": "−0.000…+0.035", "mde": "0.03 at n 180", "verdict": "not-detected", "verdict_detail": "closed: FAIL at step 0 (G2 no-harm; G1 and G3 PASS); 0 judge calls", "reopen_if": "≥ 500 pack-matched items (n for MDE 0.017), or a reader lever that converts kept chunks (GROUND-ST/STACK may carry it as a declared component)", "closed_row": "R46", "registry_public": "boar-research:reports/CLOSED.md:70", "date": "2026-10-13", "fact_ids": [ "F-1122", "F-506" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-144", "code": "DENSECHUNK-ST2", "name": "DENSECHUNK-ST2", "family": "retrieval", "question": "With dense chunk hits appended as up to 2 extra raw slots (never displacing the lexical list), does the 4B cut 7 answer the 180 public self-test items better vs packDense2 off?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "180 self-tests; v0.1 public 60 (no-harm replay); v0.2 (bench-v02) 110 no-harm replay", "judge": "Claude Opus, Claude Haiku (L1)", "n": "180", "seeds": "1 (greedy 4B, seed 42)", "result": "+0.017 (−0.000…+0.035)", "result_verified": true, "effect": "+0.017 (−0.000…+0.035)", "ci": "−0.000…+0.035", "mde": "0.03 at n 180", "verdict": "not-detected", "verdict_detail": "not decided at MDE 0.03 (Opus L2 Δ CI through 0) → parked; format v3 not shipped", "reopen_if": "≥ 500 pack-matched items (n for MDE 0.017), or a reader lever that converts kept chunks (GROUND-ST/STACK may carry it as a declared component)", "closed_row": "R46", "registry_public": "boar-research:reports/CLOSED.md:70", "date": "2026-10-13", "fact_ids": [ "F-1122", "F-507", "F-880" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-145", "code": "GROUND-ST", "name": "GROUND-ST", "family": "packs", "question": "On the 180 public self-test items, do keep-step relaxations X = packGround and X2 = packGround + keepTop1 keep more reached fact chunks/articles than OFF (published build) without kit harm (step 0 gate)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "180 self-tests; v0.1; v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "180 self-test items + v0.1 60 + v0.2 110 kit (replays, 0 judge calls)", "seeds": "n/a (deterministic replay)", "result": "chunk kept 51–54 vs bar 55", "result_verified": true, "effect": "chunk kept 51–54 vs bar 55", "ci": null, "mde": "n/a (step 0 reach gate)", "verdict": "closed", "verdict_detail": "closed: FAIL at step 0 (G1 chunk kept 51/53 vs ≥ 55). Family closed after addendum 1; X3 carried into STACK as a declared component per ledgers.", "reopen_if": "a keep rule that reaches the pack's 2nd–4th passage without broad context growth, or pack-matched items with power for a stand-alone quality test (n ≈ 500)", "closed_row": "R47", "registry_public": "boar-research:reports/CLOSED.md:71", "date": "2026-10-13", "fact_ids": [ "F-536", "F-1107" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-146", "code": "GROUND-ST addendum 1", "name": "GROUND-ST addendum 1 (X3 perpack)", "family": "packs", "question": "Does variant X3 = packGround + per-pack top-1 protection clear the unchanged step-0 gates (G1 chunk ≥ 55, article ≥ 85; G2; G3) vs OFF on the 180 self-tests?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "180 self-tests; v0.1; v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "180 self-test items + v0.1 60 + v0.2 110 kit (replays, 0 judge calls)", "seeds": "n/a (deterministic replay)", "result": "chunk kept 51–54 vs bar 55", "result_verified": true, "effect": "chunk kept 51–54 vs bar 55", "ci": null, "mde": "n/a (step 0 reach gate)", "verdict": "closed", "verdict_detail": "closed: FAIL at step 0 (G1 chunk kept 54 < 55). Family closed at step 0 (CLOSED R47); X3 carried into STACK as a declared component by coordinator decision.", "reopen_if": "a keep rule that reaches the pack's 2nd–4th passage without broad context growth, or pack-matched items with power for a stand-alone quality test (n ≈ 500)", "closed_row": "R47", "registry_public": "boar-research:reports/CLOSED.md:71", "date": "2026-10-13", "fact_ids": [ "F-536" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-147", "code": "ORACLE-ST", "name": "ORACLE-ST", "family": "retrieval (pack ceiling / reader wall)", "question": "Retrieval ceiling: does injecting the known fact chunk into the 4B cut 7 context (ORACLE-ADD extra slot; ORACLE-ONLY lead+fact chunk) on the 180 public self-tests raise the score vs the S runs (item's pack, today's build)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/oracle-st/ORACLE-ST-criterion.md" }, "set": "180 self-tests", "judge": "Claude Opus, Claude Haiku (L1)", "n": "180 (planned)", "seeds": "1 (s42, deterministic generation)", "result": "inject fact chunk (ADD) +0.095 (+0.070…+0.122); fact chunk only (ONLY) +0.115 (+0.088…+0.142)", "result_verified": true, "effect": "inject fact chunk (ADD) +0.095 (+0.070…+0.122); fact chunk only (ONLY) +0.115 (+0.088…+0.142)", "ci": "+0.070…+0.122", "mde": null, "verdict": "open", "verdict_detail": "running (stage 0 PASS; stage 1 ORACLE-ADD generation 138/180 at 21:57, ORACLE-ONLY not started, 0 judge calls; no ORACLE-ST.md yet)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-13", "fact_ids": [ "F-027" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-148", "code": "PACKROUTE", "name": "PACKROUTE", "family": "packs (retrieval routing, product configuration)", "question": "With all 9 packs + kit installed, does a deterministic vote router (top-1 pack or none; R) or a per-pack quota 2/6 (Q) beat today's all-packs-as-peers engine (P), with single-pack S as the oracle reference?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "180 self-tests; v0.1; v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "180 self-test items (+ v0.1 60 / v0.2 110 kit no-harm; 24 kit 'none' items)", "seeds": "n/a (0-LLM replays)", "result": "0.25 / 0.58 / 0.61 vs 0.85", "result_verified": true, "effect": "0.25 / 0.58 / 0.61 vs 0.85", "ci": null, "mde": "n/a (step 0)", "verdict": "closed", "verdict_detail": "closed (FAIL at step 0: R fails G1/G2/G3, Q fails G2/G3; 0 LLM / 0 judge / 0 paid)", "reopen_if": "a router ≥ 0.85 own-pack on the 180 AND ≥ 0.85 none on kit items, or a soft-route mechanism (prior in fusion, not a filter) with a declared reach gate; only if CROWDQ shows the crowding cost ≥ 0.03 raw", "closed_row": "R48", "registry_public": "boar-research:reports/CLOSED.md:72", "date": "2026-10-13", "fact_ids": [ "F-1123", "F-1106" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-149", "code": "PACKROUTE-2", "name": "PACKROUTE-2", "family": "packs (retrieval routing, product configuration)", "question": "Can the shipped 4B as a zero-shot classifier over installed pack descriptions route each question to its one relevant pack (R4; R4Q = none → peers with quota 2/6) well enough to recover single-pack reach vs peers P?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "180 self-tests; v0.1; v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "180 self-tests + 24 kit 'none' items (+ v0.1 60 / v0.2 110 distribution); ≈ 350 classifier calls per wording", "seeds": "1 (seed 42, temperature 0)", "result": "0.25 / 0.58 / 0.61 vs 0.85", "result_verified": true, "effect": "0.25 / 0.58 / 0.61 vs 0.85", "ci": null, "mde": "n/a (step 0)", "verdict": "closed", "verdict_detail": "closed (FAIL at step 0a on both wordings; 0b diagnostic run afterwards, verdict unchanged)", "reopen_if": "a router ≥ 0.85 own-pack on the 180 AND ≥ 0.85 none on kit items, or a soft-route mechanism (prior in fusion, not a filter) with a declared reach gate; only if CROWDQ shows the crowding cost ≥ 0.03 raw", "closed_row": "R48", "registry_public": "boar-research:reports/CLOSED.md:72", "date": "2026-10-13", "fact_ids": [ "F-1123", "F-1106" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-150", "code": "SEEDGEN", "name": "SEEDGEN", "family": "tooling", "question": "Is 4B self-test generation (cut 7, run_selftest.sh, MacBook CPU llama-server) byte-deterministic across repeated runs, and was DENSECHUNK-ST2's 56/102 'kept-unchanged same answer' a sampler-noise floor or a keying artefact?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "180 self-tests (qa-health-se 20-item pack; ST2 re-analysis over 9 packs)", "judge": "none (deterministic gate)", "n": "20 requests (qa-health-se OFF, 15 LLM calls) × 4 runs; ST2 re-analysis 102 kept-unchanged items over 9 packs", "seeds": "1 (seed 42 on every request)", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling (done; generation byte-deterministic on the MacBook CPU path; RULE 27)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-13", "fact_ids": [ "F-881" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-151", "code": "SELFTEST-DEEP", "name": "SELFTEST-DEEP", "family": "deep-tier / packs", "question": "How much does each published pack lift the deep tier H (dt512: Qwen3.6-35B-A3B + thinking 512) on its own 20 self-test items, vs H without the pack, and vs the 4B with the pack?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/packs/SELFTEST-DEEP-criterion.md" }, "set": "180 self-tests", "judge": "Claude Haiku (L1)", "n": "180 (9 packs × 20); 18 runs; 720 Haiku calls", "seeds": "1 (42)", "result": "deep with vs without pack Δ −0.002 (−0.020…+0.016); deep − 4B with the same pack +0.054 (+0.034…+0.073)", "result_verified": true, "effect": "deep with vs without pack Δ −0.002 (−0.020…+0.016); deep − 4B with the same pack +0.054 (+0.034…+0.073)", "ci": "−0.020…+0.016", "mde": null, "verdict": "measurement", "verdict_detail": "measurement (done 2026-10-13 19:36; 18/18 runs; no verdict rule)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-13", "fact_ids": [ "F-029", "F-1052" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-152", "code": "SELFTEST-SCORES", "name": "SELFTEST-SCORES", "family": "packs", "question": "On each published pack's 20 public self-test items, how much does adding the pack to the 4B claim build (cut 7 + claim kit) lift the score vs claim kit only?", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "180 self-tests", "judge": "Claude Haiku (L1)", "n": "180 (9 packs × 20); wave-3 addendum 40", "seeds": "1 (42)", "result": "Opus +0.042 (+0.024…+0.062); Haiku +0.032 (+0.016…+0.050)", "result_verified": true, "effect": "Opus +0.042 (+0.024…+0.062); Haiku +0.032 (+0.016…+0.050)", "ci": "+0.024…+0.062", "mde": null, "verdict": "measurement", "verdict_detail": "measurement (done 2026-10-13 10:37; no verdict rule)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-13", "fact_ids": [ "F-025" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-153", "code": "SFT4B", "name": "SFT4B", "family": "data/finetune", "question": "Can a small LoRA (rank 16, ~1,200 extraction-style grounded QA examples from the packs' own text, filtered by jev quote_supports >= 0.9) teach the claim 4B reader to use the kept pack passage, measured as X (claim 4B + LoRA) vs C (claim 4B, SELFTEST-SCORES with-pack traces) on the 180 pack-matched self-tests, with v0.1 public 60 no-harm.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "180 self-tests; v0.1", "judge": "none (deterministic gate)", "n": "180 (eval, planned); 1,200 train / 100 hold-out examples built", "seeds": "1 (deterministic generation; training seed 42)", "result": "pass 38 %, 562 kept vs ≥ 1,000", "result_verified": true, "effect": "pass 38 %, 562 kept vs ≥ 1,000", "ci": null, "mde": "n/a (stage 0)", "verdict": "closed", "verdict_detail": "open / in progress: stage 0 data PASS (v1, then v2); stage 1 environment + parity tooling PASS on the smoke adapter (STAGE1.md); training (step 6) on hold until v3 data (addendum 2); no verdict", "reopen_if": "the deep-tier teacher was probed 10-14 (120 items): 6/120 pass vs Haiku 54/120 → not a stronger teacher as run; remaining: an Opus teacher (≥ 50 % round-1 at quote_supports ≥ 0.9) or a 0.8 bar validated by the strict 20-item review ≤ 6/20", "closed_row": "R71", "registry_public": "boar-research:reports/CLOSED.md:100", "date": "2026-10-13", "fact_ids": [ "F-549", "F-508", "F-640" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-154", "code": "STACK", "name": "STACK", "family": "packs", "question": "Step-0 replay gate: does STACK-lite (packGround 'perpack' keep rule X3 + packRoute 'quota' 2/6) under all 9 packs + kit raise fact article/chunk reach vs P (CROWDQ peers, cut 7, all 9 installed) without kit harm or latency cost.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/stack/STACK-criterion.md" }, "set": "180 self-tests; v0.1; v0.2 (bench-v02)", "judge": "none (deterministic gate)", "n": "180 self-tests (+ v0.1 60, v0.2 110 kit replays)", "seeds": "1 (deterministic replay)", "result": "+0.005 (−0.009…+0.021)", "result_verified": true, "effect": "+0.005 (−0.009…+0.021)", "ci": "−0.009…+0.021", "mde": "0.03 at n 180", "verdict": "not-detected", "verdict_detail": "closed (FAIL at step 0 on the G1 fact-chunk bar 36 < 39; measurement continued as STACK-2; pack stack closed in R49)", "reopen_if": "≥ 500 pack-matched items or an SE-only product decision (SE Δ CI low −0.000)", "closed_row": "R49", "registry_public": "boar-research:reports/CLOSED.md:73", "date": "2026-10-13", "fact_ids": [ "F-1124", "F-1108" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-155", "code": "STACK-2", "name": "STACK-2", "family": "packs", "question": "Quality measurement after STACK's step-0 FAIL: X (packGround 'perpack' + packRoute 'quota', all 9 packs + kit, cut 7) vs P (CROWDQ peers run) on the 180 self-tests, plus v0.1 public 60 no-harm ON vs OFF (kit only).", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/stack/STACK2-criterion.md" }, "set": "180 self-tests; v0.1", "judge": "Claude Haiku (L1)", "n": "180 (+ v0.1 57 judged)", "seeds": "1 (deterministic, seed 42)", "result": "+0.005 (−0.009…+0.021)", "result_verified": true, "effect": "+0.005 (−0.009…+0.021)", "ci": "−0.009…+0.021", "mde": "0.03 at n 180", "verdict": "not-detected", "verdict_detail": "closed: not decided at MDE 0.03 → pack stack closed (R49); optional setting, not default", "reopen_if": "≥ 500 pack-matched items or an SE-only product decision (SE Δ CI low −0.000)", "closed_row": "R49", "registry_public": "boar-research:reports/CLOSED.md:73", "date": "2026-10-13", "fact_ids": [ "F-1124", "F-1047", "F-1108" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-156", "code": "T-K6b", "name": "T-K6b", "family": "packs (retrieval / coverage)", "question": "Crowding and V-para coverage probes for the 2 wave-2 packs (boar-factbook, boar-simplewiki-en) under the frozen share ≤ 0.40 / kit-gold-not-lower rule.", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "v0.1; v0.2 (bench-v02); V-para (V-para)", "judge": "none (deterministic gate)", "n": "v0.2 110 + v0.1 60 (crowding); 34 V-para absent items (coverage)", "seeds": "n/a", "result": "n/a", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "closed (probe: boar-simplewiki-en user-selected only; boar-factbook user-selected only by coordinator call, frozen aggregator said eligible)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-13", "fact_ids": [], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-157", "code": "VITALPACK-ADD", "name": "VITALPACK-ADD", "family": "packs", "question": "Does appending one wiki-vital5 intro after grounding (additive quota k=1, pool 6, never displacing curated chunks) to the deep thinking tier move its score on v0.2 (direction) and V-para (confirmation), vs control H (dt512 rows; TV twin for the post-hoc Haiku full screen)?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "v0.1 (smoke); v0.2 (bench-v02) random20 slice + full 110; V-para (generated only)", "judge": "Claude Opus, Claude Haiku (L1)", "n": "smoke 3 v0.1; slice 20; v0.2 full 110; V-para s42 58 generated, unjudged", "seeds": "1 (s42); s43 never generated", "result": "+0.0012 (−0.012…+0.015)", "result_verified": true, "effect": "+0.0012 (−0.012…+0.015)", "ci": "−0.012…+0.015", "mde": "0.025 (v0.2, 1 seed); the primary's MDE 0.03 (V-para, 2 seeds) never measured", "verdict": "not-detected", "verdict_detail": "closed in practice: NO at the slice (Opus −0.011) and not detected at L2 on v0.2 full 110 (Opus +0.0012, CI −0.0120…+0.0147); frozen stage 4 (V-para 2 seeds) never run, so no frozen LEVER/SMALL/NO verdict", "reopen_if": "stage 4 (V-para s42 + s43, 204 Opus) only if s43 A traces are ever generated (VITALADD.md:19)", "closed_row": "R68", "registry_public": "boar-research:reports/CLOSED.md:97", "date": "2026-10-13", "fact_ids": [ "F-256" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-158", "code": "WAVE3", "name": "WAVE3", "family": "packs", "question": "Build, probe (crowding, V-para coverage), self-test score and publish two wave-3 Stack Exchange packs (qa-cooking-se, qa-money-se); self-test with vs without the pack on the 4B.", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "wave-3 self-tests; v0.1; v0.2 (bench-v02); V-para (V-para)", "judge": "Claude Haiku (L1)", "n": "40 self-test items (20 per pack); v0.1 60 / v0.2 110 crowding; 34 V-para absent items", "seeds": "1 (42; draw seed 11)", "result": "Haiku: cooking +0.010 (−0.039…+0.065); money +0.030 (−0.007…+0.072); pooled 40 +0.020 (−0.011…+0.054)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "done (product: qa-cooking-se default-eligible, qa-money-se user-selected only; published on HF; no lever verdict)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-13", "fact_ids": [], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-159", "code": "WIKIFULL-REOPEN", "name": "WIKIFULL-REOPEN", "family": "retrieval", "question": "Pooled stage A (15 FineWiki shards) + stage B (enwiki 20261001 pages outside FineWiki): is the deep tier's absent V-para fact anywhere in English Wikipedia, deciding REOPEN ≥ 14 / MARGINAL 7–13 / CLOSE ≤ 6 for Wikipedia as kit content.", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "V-para (V-para); v0.2 (bench-v02, dev)", "judge": "none (deterministic gate)", "n": "34 V-para (78 v0.2 dev context)", "seeds": "n/a (jev deterministic)", "result": "9/34 verified", "result_verified": true, "effect": "9/34 verified", "ci": null, "mde": "n/a (0-call gate)", "verdict": "closed", "verdict_detail": "closed (frozen band MARGINAL, 9/34 verified, lower bound; no WIKI-ALL arm; Wikipedia neither reopened nor closed as kit content per the frozen wording)", "reopen_if": "a question-shaped or domain corpus that covers the comparison/explanation/synthesis misses (T-K4d: the wave-1 packs cover 1/34); or pack-matched items where coverage is by construction", "closed_row": "R45", "registry_public": "boar-research:reports/CLOSED.md:69", "date": "2026-10-13", "fact_ids": [ "F-1121" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-160", "code": "H8-ST", "name": "Contextual chunk headers in the lexical index", "family": "4B (retrieval)", "question": "Contextual chunk headers in the lexical index (H8-ST step 0, H8 (a); article first sentence as a 4th FTS column)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (129b8e22…)", "where_public": null }, "set": "180 public self-tests + v0.1 60 (kit), 11 published packs, OFF byte-identical stage 0 (180/180, 60/60); 0 LLM, 0 judge, 0 paid; criterion 129b8e22…", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "top-8 69 → 75 (bar 90)", "result_verified": true, "effect": "top-8 69 → 75 (bar 90)", "ci": null, "mde": "**FAIL at step 0 (0 judge calls): G1 (top-8 75 < 90; kept +2 < +15) and G2 (gold-kept 8 → 7) fail, G3 passes**; the lever H8 (a) assumed does not exist in this engine. Reopen: question-shaped text per chunk (H8 (b) doc2query, ≈ 14 h of 4B generation per 55k chunks on the mini) or a dense sidecar that does not displace (R46); untested: lead weight 2, lead in the section column, one merged column", "verdict": "closed", "verdict_detail": "arms/h8st/H8ST.md", "reopen_if": null, "closed_row": "R86", "registry_public": null, "date": "2026-10-13", "fact_ids": [ "F-426", "F-967" ], "chapter": "walls", "source": "closed registry row R86 + verified fact" }, { "id": "X-161", "code": "HEADPASS-X2b", "name": "HEADPASS-X2b: neighbour chunks + budgeted dense candidates appended after the keep step", "family": null, "question": "Does appending the retrieved article's neighbouring chunks plus up to 2 dense-vector candidates (≤ +600 prompt tokens) raise the 4B's score on the 180 pack-matched self-tests?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": "boar-research:arms/headpass-st/HEADPASS-ST.md:89" }, "set": "180 pack self-tests (dev)", "judge": "Claude Opus (L2) and Claude Haiku (L1)", "n": null, "seeds": null, "result": "Opus +0.066 (+0.043…+0.091); Haiku +0.056; fact chunk present in context 35 → 102 of 180", "result_verified": true, "effect": "Opus +0.066 (+0.043…+0.091); Haiku +0.056; fact chunk present in context 35 → 102 of 180", "ci": "+0.043…+0.091", "mde": null, "verdict": "adopted", "verdict_detail": null, "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-14", "fact_ids": [ "F-028", "F-1069" ], "chapter": "packs", "source": "canonical facts" }, { "id": "X-162", "code": "MICROD2Q", "name": "Build-time query expansion", "family": "4B (retrieval, index)", "question": "Build-time query expansion (doc2query) for the pack line (MICROD2Q: synthetic 4B questions on the 137 fact chunks + their articles + top-39 competitors; index-only column and gate-aware variant)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (2bca6998…)", "where_public": null }, "set": "180 self-tests, replays only; 5,975 chunks generated on the MacBook (Metal); 0 judge calls; v0.1 no-harm", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "reach 69 → 77 vs gate 90", "result_verified": true, "effect": "reach 69 → 77 vs gate 90", "ci": null, "mde": "n/a (step 0)", "verdict": "closed", "verdict_detail": "FAIL at step 0 (side session, frozen criterion 2bca6998…)", "reopen_if": "a query writer that reaches the gate/keep misses (the 56 wrong-chunk + 20 keep + 21 gate items) rather than the 27 lexical misses; no GPU build bought", "closed_row": "R72", "registry_public": "boar-research:reports/CLOSED.md:101", "date": "2026-10-14", "fact_ids": [ "F-550", "F-968" ], "chapter": "packs", "source": "closed registry row R72 + verified fact" }, { "id": "X-163", "code": "HEADPASS-X2", "name": "Dense chunk candidates appended unbudgeted through the post-grounding append path", "family": "4B (retrieval → context)", "question": "Dense chunk candidates appended unbudgeted through the post-grounding append path (HEADPASS-X2: ST2 dense top-2 + X1 neighbours, no displacement)", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": "boar-research:arms/headpass-st/HEADPASS-X2-criterion.md" }, "set": "180 self-tests replays + v0.1/v0.2 kit no-harm; 0 judge calls", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "reach PASS but +20.5 tokens over cost gate", "result_verified": true, "effect": "reach PASS but +20.5 tokens over cost gate", "ci": null, "mde": "n/a (step 0)", "verdict": "closed", "verdict_detail": "FAIL at step 0 on the prompt-token gate (by 20.5 tokens; frozen rule)", "reopen_if": "resolved by X2b (budget-capped append, addendum 1): LEVER +0.066 (+0.043…+0.091) Opus vs S on the 180 (10-14 17:1x); X3 (reranker-chosen slots) running", "closed_row": "R73", "registry_public": "boar-research:reports/CLOSED.md:102", "date": "2026-10-14", "fact_ids": [ "F-1126", "F-1109" ], "chapter": "packs", "source": "closed registry row R73 + verified fact" }, { "id": "X-164", "code": "HEADPASS-X3", "name": "Cross-encoder reranker choosing the appended chunks on the HEADPASS path", "family": "4B (retrieval → context)", "question": "Cross-encoder reranker choosing the appended chunks on the HEADPASS path (X3a reranker-only, X3b X1 + reranker; bge-reranker-v2-m3 over BM25 ∪ dense ∪ top-article candidates, under the X2b token budget)", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": "boar-research:arms/headpass-st/HEADPASS-X3-criterion.md" }, "set": "180 self-tests replays + v0.1/v0.2 kit no-harm; 0 judge calls; reranker latency timed", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "reach 111/109 vs 117; 21.4 s/item", "result_verified": true, "effect": "reach 111/109 vs 117; 21.4 s/item", "ci": null, "mde": "n/a (step 0)", "verdict": "closed", "verdict_detail": "FAIL at step 0 (reach below the power bar; device cost rules out a default)", "reopen_if": "a reranker ≤ 1 s/item on device with the health-strict rule applied before ranking; M2 (pack-agnostic rerank under many packs: article kept 41 → 64, 3.9 s/item, 9 emptied contexts) parked as a declared variant", "closed_row": "R74", "registry_public": "boar-research:reports/CLOSED.md:103", "date": "2026-10-14", "fact_ids": [ "F-551", "F-1110" ], "chapter": "packs", "source": "closed registry row R74 + verified fact" }, { "id": "X-165", "code": "DEVICE-X6", "name": "Cut 7 on a 12 GB-class phone", "family": "4B (device)", "question": "Cut 7 on a 12 GB-class phone (DEVICE-X6, measurement; a POCO X6 Pro test phone, MT6897, 10.7 GiB)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (43ec8428…)", "where_public": "boar-research:arms/device-x6/DEVICE-X6-criterion.md" }, "set": "v0.1 public 57 ids, Haiku L1 N = 4 both orders, 114 phone-run calls, desktop rows reused; 0 paid; criterion 43ec8428… before any call", "judge": "Claude Haiku (L1)", "n": null, "seeds": null, "result": "phone = desktop", "result_verified": true, "effect": "phone = desktop", "ci": null, "mde": "MEASURED: device = desktop within noise (declared reading); not a lever. Reopens nothing; enables the deep-tier device steps (D2/D3/D5)", "verdict": "closed", "verdict_detail": "arms/device-x6/DEVICE-X6.md", "reopen_if": "", "closed_row": "R75", "registry_public": "boar-research:reports/CLOSED.md:104", "date": "2026-10-14", "fact_ids": [ "F-817", "F-023" ], "chapter": "phone", "source": "closed registry row R75 + verified fact" }, { "id": "X-166", "code": "LONGSESSION", "name": "Multi-turn drift on the 4B", "family": "4B (history → context)", "question": "Multi-turn drift on the 4B (LONGSESSION, measurement; the bounty issuer's criterion #1)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (710f07a6…)", "where_public": "boar-research:arms/longsession/LONGSESSION-criterion.md" }, "set": "50 agent-written dev questions (10×5), paired in-session vs standalone, Haiku N = 4 both orders, 102 calls; 0 paid; criterion 710f07a6… before any run", "judge": "Claude Haiku (L1)", "n": null, "seeds": null, "result": "no drop to turn 15", "result_verified": true, "effect": "no drop to turn 15", "ci": null, "mde": "MEASURED: no drop in 5 turns nor out to turn 15 (ext15: blocks 1.42 / 1.10 / 1.18, none below 1); history helps on follow-ups; cost = latency (≈ 22 s vs 8 s at turns 14–15, desktop); app truncation gap turns 5–7 noted; not a lever", "verdict": "closed", "verdict_detail": "arms/longsession/LONGSESSION.md", "reopen_if": "", "closed_row": "R76", "registry_public": "boar-research:reports/CLOSED.md:105", "date": "2026-10-14", "fact_ids": [ "F-552", "F-030" ], "chapter": "handoff", "source": "closed registry row R76 + verified fact" }, { "id": "X-167", "code": "BONSAI2-9B", "name": "Bonsai 2 ternary 9B", "family": "4B tier (8 GB)", "question": "Bonsai 2 ternary 9B (Qwen3.5-9B PTQ1_0, 2.0 GB) as a faster/smaller fast tier (BONSAI2 S0 + kernel sweep; NEWMODEL family)", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": null }, "set": "mini M4 CPU, llama-bench pp512/tg64 + 5 public v0.1 items; PrismML fork `prism-b10777-e3de715` generic + kleidiai binaries, -t 4/8/10, -fa 1; 0 judge calls, 0 paid; pre-freeze smoke (S0)", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "decode 0.49× the 4B (bar 0.7×)", "result_verified": true, "effect": "decode 0.49× the 4B (bar 0.7×)", "ci": null, "mde": "**closed at S0 on the speed clause (≥ 0.7× the 4B; PLAN reading rule)**; quality never measured (no S1). Reopen: PrismML ships faster CPU kernels for PTQ1_0/PQ2_0, or the native app gets a GPU path (the company's phone claims are GPU). 27B parked for disk (mini at the 25 GiB floor; fits only by deleting the 35B deep model = the author). Model deleted (class A; recipe in MODELS.md)", "verdict": "closed", "verdict_detail": "arms/bonsai2/S0.md", "reopen_if": null, "closed_row": "R77", "registry_public": null, "date": "2026-10-15", "fact_ids": [ "F-553" ], "chapter": "walls", "source": "closed registry row R77 + verified fact" }, { "id": "X-168", "code": "BONSAI2-4B", "name": "Bonsai 2 ternary 4B", "family": "4B tier", "question": "Bonsai 2 ternary 4B (Qwen3.5-4B PTQ1_0, 0.95 GB) as a fast-tier swap for the Qwen3-4B Q4_K_M (BONSAI2 S1, parity question; the author 10-15 asked for the 4B and the 2B)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (aadebf08…)", "where_public": null }, "set": "random20-v02 slice (dev, scoped exception), Haiku 5.5-low N = 4 both orders, 44 calls (40 + 4 preflight), 0 Opus, 0 paid; criterion aadebf08… frozen before any run; smoke PASS (0 reasoning tokens wit…", "judge": "Claude Opus, Claude Haiku (L1)", "n": null, "seeds": null, "result": "−0.045 (−0.109…+0.004)", "result_verified": true, "effect": "−0.045 (−0.109…+0.004)", "ci": "−0.109…+0.004", "mde": "**NO (Haiku, dev slice): CI low < −0.05 under the frozen parity clause**; not a near-miss (Δ 0.015 past −0.03). Prior held (Q35 base −0.05). 2B (0.43 GB, 1.7× faster, 0.9 GB RSS) = lite-tier product idea only, with the danger cards in front (inverted boil/chlorinate advice unjudged on S0). 27B parked (disk). Both models deleted (class A; recipes in MODELS.md); fork binary kept. Reopen: PrismML faster CPU kernels AND a stronger base, or a device tier where 0.5× RSS decides", "verdict": "closed", "verdict_detail": "arms/bonsai2/S1.md", "reopen_if": null, "closed_row": "R78", "registry_public": null, "date": "2026-10-15", "fact_ids": [ "F-554" ], "chapter": "walls", "source": "closed registry row R78 + verified fact" }, { "id": "X-169", "code": "MATHROUTE", "name": "A3 router", "family": "4B (cut 7 routing)", "question": "A3 router (skip retrieval + calculator off) as a substitute for the calculator on math/unit questions (MATHROUTE, tag probe, measurement arm)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (e0124ac5…)", "where_public": null }, "set": "mathroute-v1 slice 26 EN items (2 v0.1 calculator items + 24 agent-written dev), arms A cut 7 / B A3 / C no calc + retrieval, seed 42, deterministic scorer; 0 judge, 0 paid; criterion e0124ac5… froze…", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "20/26 vs 23/26", "result_verified": true, "effect": "20/26 vs 23/26", "ci": null, "mde": "**closed: A3 does not recover the calculator's advantage** (B vs A −3, n 26, direction only). Reopen: only via the numeric guard before the health refusal (→ NUMGUARD), not the router; A3 stays a possible add-on for model-answered math (no retrieval → no fake [n])", "verdict": "closed", "verdict_detail": "arms/mathroute/MATHROUTE.md", "reopen_if": null, "closed_row": "R80", "registry_public": null, "date": "2026-10-15", "fact_ids": [ "F-424", "F-382" ], "chapter": "builds", "source": "closed registry row R80 + verified fact" }, { "id": "X-170", "code": "NUMGUARD", "name": "Lexicon numeric guard before the detector's health refusal", "family": "4B (routing, deterministic)", "question": "Lexicon numeric guard before the detector's health refusal (NUMGUARD r1 → NUMGUARD-2 → r2 → FINAL, confirmation REVIEW; r3 calculator-only is the surviving arm, NOT closed)", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": null }, "set": "mathroute-v1 26 + v0.1 public 60 + 180 self-tests + 32/6/18/2 health and multi-turn probes, seed 42; 0 judge, 0 paid; criteria b867a27c… / d79b7919… / e7231bf4… / a63c928e… frozen before each run", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "review: 8 of 20 new probes went to model closed-book", "result_verified": true, "effect": "review: 8 of 20 new probes went to model closed-book", "ci": null, "mde": "**closed as unsafe after the confirmation review (REJECT, 4 must-fix): a lexicon guard that lets the model answer a detector-fired turn cannot be made safe**; passed gates were tuned on the same probe set. Surviving arm = NUMGUARD-R3 (calculator-only, a model never answers a fired turn; PASS 7/7, slice 24/26, mr-a04 recovered, mr-u07/mr-a02 back to refusals; patch shipped to PR #85). Reopen: pressure (bar/psi) and clock-add calculators (follow-up after 10-20); any new calculator must re-run the health goldens", "verdict": "closed", "verdict_detail": "arms/numguard/NUMGUARD.md, NUMGUARD2.md, NUMGUARD-R2.md, NUMGUARD-FINAL.md, REVIEW-numguard-final.md; surviving: NUMGUARD-R3.md", "reopen_if": null, "closed_row": "R81", "registry_public": null, "date": "2026-10-15", "fact_ids": [ "F-425", "F-386", "F-417", "F-419" ], "chapter": "builds", "source": "closed registry row R81 + verified fact" }, { "id": "X-171", "code": "BONSAI2-2B", "name": "Bonsai 2 ternary 2B", "family": "lite (2B)", "question": "Bonsai 2 ternary 2B (Ternary-Bonsai-2-2B PTQ1_0, 0.43 GB) as the lite fast model behind cut 7 (BONSAI2-2B; Q1 kit slice + Q2 health layers)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (9f2757bb…)", "where_public": null }, "set": "Q1 random20-v02 (dev, scoped exception) kit, Haiku 5.5-low N = 4 both orders vs the 4B dv-C rows; Q2 64 health/danger items full engine, both models; 44 Haiku, 0 Opus, 30 jev calls; criterion…", "judge": "Claude Opus, Claude Haiku (L1), calibrated classifier (jev)", "n": null, "seeds": null, "result": "−0.168; 6/15 unsafe/inverted model-only health answers", "result_verified": true, "effect": "−0.168; 6/15 unsafe/inverted model-only health answers", "ci": null, "mde": "**measurement, no verdict label; the 2B is not a quality substitute (CI excludes −0.10)**. Real finding: health coverage gap, the 15 model-only items are exactly those the detector does not fire on. Reopen: a lite tier needs a health-shaped predicate (drug, dose, toxic, CO, survival) switching on a refusal or warning before it faces users (→ WARNSPEC); dose and arithmetic questions to calculator or card", "verdict": "closed", "verdict_detail": "arms/bonsai2/BONSAI2-2B.md", "reopen_if": null, "closed_row": "R82", "registry_public": null, "date": "2026-10-15", "fact_ids": [ "F-555", "F-791" ], "chapter": "builds", "source": "closed registry row R82 + verified fact" }, { "id": "X-172", "code": "LITE2", "name": "Ternary 2B vs stock Qwen3-1.7B Q4_K_M as the lite fast model", "family": "lite (1.7B / 2B)", "question": "Ternary 2B vs stock Qwen3-1.7B Q4_K_M as the lite fast model (LITE2)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (e2c88e76…)", "where_public": null }, "set": "random20-v02 kit + selftest20 (2 per pack over the 9 packs of the 180, +2), Haiku 5.5-low N = 4 both orders, 120 Haiku, 0 Opus, 0 paid; criterion e2c88e76… frozen after the 5-item smoke, before any j…", "judge": "Claude Opus, Claude Haiku (L1)", "n": null, "seeds": null, "result": "−0.057 (−0.104…−0.014) kit", "result_verified": true, "effect": "−0.057 (−0.104…−0.014) kit", "ci": "−0.104…−0.014", "mde": "**LOSS (Haiku, dev slices): ternary 2B does not match the stock-runtime 1.7B**; the 2B keeps only its footprint edge. 1.7B = the lite candidate (LITE3/LITE4). Reopen: a ternary model that closes the quality gap, or a 4–6 GB device where 3× less RAM decides *(review CRITIQUE-2026-10-16: not detected at MDE ≈ 0.08 (n ≈ 20))*", "verdict": "not-detected", "verdict_detail": "arms/lite2/LITE2.md | relabelled per the 2026-10-16 outside critique: not detected at MDE 0.08 (n ≈ 20)", "reopen_if": null, "closed_row": "R83", "registry_public": null, "date": "2026-10-15", "fact_ids": [ "F-1302", "F-1075" ], "chapter": "handoff", "source": "closed registry row R83 + verified fact" }, { "id": "X-173", "code": "LITE4", "name": "Source-bound reading rule on the lite 1.7B", "family": "lite 1.7B", "question": "Source-bound reading rule on the lite 1.7B (deps.liteSourceBound; \"one hedged sentence\" on memory/nosource routes; LITE4)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (6cab3a8a…)", "where_public": null }, "set": "random20-v02 (dev, scoped exception), Haiku 5.5-low N = 4 both orders, 40 Haiku, 0 Opus, jev 0 (traces hold no passage text); 0-call gate G0–G5 PASS (10/10 invented answers changed, +38 prompt tokens…", "judge": "Claude Opus, Claude Haiku (L1), calibrated classifier (jev)", "n": null, "seeds": null, "result": "−0.001 (−0.064…+0.063)", "result_verified": true, "effect": "−0.001 (−0.064…+0.063)", "ci": "−0.064…+0.063", "mde": "**DIRECTION-NEGATIVE by the frozen rule (Δ < 0); read as flat, inside judge noise; not worth a larger-n run**. Reopen: a new arm with a new freeze that keeps the source-bound line for sourced answers and drops the one-sentence cap on memory/nosource, or attacks the \"thin\" bucket (3–4 sentence floor) *(review CRITIQUE-2026-10-16: not detected at MDE ≈ 0.08 (n ≈ 20))*", "verdict": "not-detected", "verdict_detail": "arms/lite4/LITE4.md | relabelled per the 2026-10-16 outside critique: not detected at MDE 0.08 (n ≈ 20)", "reopen_if": null, "closed_row": "R84", "registry_public": null, "date": "2026-10-15", "fact_ids": [ "F-257", "F-1298" ], "chapter": "handoff", "source": "closed registry row R84 + verified fact" }, { "id": "X-174", "code": "LESPORT-L1", "name": "LESPORT / LESPORT-L1: expert-streaming fork for the 35B", "family": null, "question": "Can an expert-streaming llama.cpp fork run the 35B deep model in a small RAM budget with judged-equivalent answers?", "hypothesis": null, "preregistered": { "status": "see arm report", "where_public": null }, "set": "v0.1 R450 contexts (public dev)", "judge": "Claude Opus", "n": null, "seeds": null, "result": "judged 0.625 vs llama-server 0.634, paired −0.010 (−0.042…+0.018); RSS 2.75–2.81 GiB; 14–17 tok/s on the Mac mini CPU; 10/10 greedy-identical vs the fork base", "result_verified": true, "effect": "judged 0.625 vs llama-server 0.634, paired −0.010 (−0.042…+0.018); RSS 2.75–2.81 GiB; 14–17 tok/s on the Mac mini CPU; 10/10 greedy-identical vs the fork base", "ci": "−0.042…+0.018", "mde": null, "verdict": "parity", "verdict_detail": null, "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-036" ], "chapter": "phone", "source": "canonical facts" }, { "id": "X-175", "code": "XFAM", "name": "XFAM: cross-family re-judge of the existing V-para traces", "family": null, "question": "Does a judge from another model family change the levels, ranking or win-rate of the 4B and deep-tier rows on V-para?", "hypothesis": null, "preregistered": { "status": "no (post-hoc)", "where_public": null }, "set": "V-para (dev-style)", "judge": "cross-family (OpenAI gpt-5.6-sol (via the Codex CLI))", "n": null, "seeds": null, "result": "4B: raw 0.497 (0.449–0.547), floor-adjusted 0.361; deep + think 512: raw 0.667 (0.621–0.711), floor-adjusted 0.577 (0.519–0.632); deep − 4B +0.171 (+0.123…+0.219) vs +0.139 under Claude; per-item Spearman 0.88–0.92; win-rate 0.00–0.01", "result_verified": true, "effect": "4B: raw 0.497 (0.449–0.547), floor-adjusted 0.361; deep + think 512: raw 0.667 (0.621–0.711), floor-adjusted 0.577 (0.519–0.632); deep − 4B +0.171 (+0.123…+0.219) vs +0.139 under Claude; per-item Spearman 0.88–0.92; win-rate 0.00–0.01", "ci": "0.449…0.547", "mde": null, "verdict": "measurement", "verdict_detail": null, "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-017", "F-1218" ], "chapter": "measurer", "source": "canonical facts" }, { "id": "X-176", "code": "LITE5-180", "name": "Pack retrieval reach for the lite tier", "family": "lite 1.7B", "question": "Pack retrieval reach for the lite tier (HEADPASS-X2b on Qwen3-1.7B Q4_K_M; LITE5 → LITE5-180)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (7fb8926d…)", "where_public": null }, "set": "180 public self-tests (dev), 1 seed, both arms regenerated on the MacBook, Haiku 5.5-low N = 4 both orders, 720 calls, 0 Opus, 0 paid; criterion 7fb8926d… frozen before any run", "judge": "Claude Opus, Claude Haiku (L1)", "n": null, "seeds": null, "result": "+0.071 (+0.048…+0.094)", "result_verified": true, "effect": "+0.071 (+0.048…+0.094)", "ci": "+0.048…+0.094", "mde": "≈ 0.03", "verdict": "adopted", "verdict_detail": "**LEVER (Haiku, dev 180)** (coordinator to review)", "reopen_if": "Opus pass or phone prefill cost if it goes on a card", "closed_row": "R79", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-1127", "F-1079", "F-1087" ], "chapter": "packs", "source": "closed registry row R79 + verified fact" }, { "id": "X-177", "code": "LITE7", "name": "Lite-only \"3-4 sentences + each figure\" style swap on the 1.7B + X2b", "family": "lite 1.7B", "question": "Lite-only \"3-4 sentences + each figure\" style swap on the 1.7B + X2b (LITE7 v2; v1 stopped at its 0-call gate on a trailing-citation artefact)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (8d252058…)", "where_public": null }, "set": "LITE5-180 ON rows reused as OFF; ON regenerated on the 180 public self-tests (MacBook, seed 42); Haiku 5.5-low N = 4 both orders on random20 seed 616, 40 calls, 0 Opus, 0 paid; criterion v2 8d252058……", "judge": "Claude Opus, Claude Haiku (L1)", "n": null, "seeds": null, "result": "+0.016 (−0.035…+0.066)", "result_verified": true, "effect": "+0.016 (−0.035…+0.066)", "ci": "−0.035…+0.066", "mde": "**DIRECTION-POSITIVE by the frozen rule, read as flat**: does not meet the 180 licence (Δ ≥ +0.03 and CI low > 0); length is not the lite model's \"thin\" problem. Reopen: fact selection (not_read, e.g. a \"key fact first\" line) or the deep/4B tier; a longer style line is not worth a 360-call run *(review CRITIQUE-2026-10-16: not detected at MDE ≈ 0.08 (n ≈ 20))*", "verdict": "not-detected", "verdict_detail": "arms/lite7/LITE7.md | relabelled per the 2026-10-16 outside critique: not detected at MDE 0.08 (n ≈ 20)", "reopen_if": null, "closed_row": "R85", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-258", "F-969" ], "chapter": "packs", "source": "closed registry row R85 + verified fact" }, { "id": "X-178", "code": "LITE8", "name": "Fact selection for the lite tier", "family": "lite 1.7B", "question": "Fact selection for the lite tier (LITE8: the not_read bucket of 1.7B + X2b vs 4B + X2b)", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": null }, "set": "0 calls; ceiling arithmetic from LITE6 (not_read = 22 % of the −0.065 gap on the 180)", "judge": null, "n": null, "seeds": null, "result": "≈ +0.014", "result_verified": true, "effect": "≈ +0.014", "ci": null, "mde": "**closed at S2: ceiling below MDE 0.03 (RULE 29)**; dedupe: SECTIONHIT R70 and GROUND-ST R47 cover selection rules on the 4B. Reopen: a selection mechanism with a measured reach gain ≥ 0.05 on the lite model, or a larger not_read bucket on an unseen set", "verdict": "closed", "verdict_detail": "arms/lite6/LITE6.md, BACKLOG row 3", "reopen_if": null, "closed_row": "R87", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-259", "F-1299" ], "chapter": "packs", "source": "closed registry row R87 + verified fact" }, { "id": "X-179", "code": "TK2-FIX", "name": "Pack sources fix v2 on the 4B", "family": "4B (retrieval)", "question": "Pack sources fix v2 on the 4B (TK2-FIX: bare title for the T-K2 sources + title lookup for codes 10–12)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (9a650fff…)", "where_public": null }, "set": "180 public self-tests, 1 seed, Haiku 5.5-low N = 4 both orders on ON (360 calls), OFF = SELFTEST-SCORES rows; 0 Opus, 0 paid; criterion 9a650fff…", "judge": "Claude Opus, Claude Haiku (L1)", "n": null, "seeds": null, "result": "+0.010 (−0.003…+0.024)", "result_verified": true, "effect": "+0.010 (−0.003…+0.024)", "ci": "−0.003…+0.024", "mde": "**NO (Haiku, dev 180)** by rule (expected_delta < MDE 0.03, pre-declared); not a harm: mechanism positive where it acts, so the arm closes as a score lever, not as a fix (ship call stays with the coordinator). Reopen: a set where the T-K2 sources hold the fact on ≥ 15 % of items", "verdict": "closed", "verdict_detail": "arms/tk2fix/TK2FIX-L1.md, TK2FIX-STEP0.md, BACKLOG row 4", "reopen_if": null, "closed_row": "R88", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-1128", "F-970" ], "chapter": "walls", "source": "closed registry row R88 + verified fact" }, { "id": "X-180", "code": "HEALTHFALLBACK", "name": "Health no-source fallback from memory", "family": "deep (health path)", "question": "Health no-source fallback from memory (HEALTHFALLBACK / H7: deep model answers a health-strict turn with no source, under a mandatory safety line)", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": null }, "set": "0 calls; re-scope over existing traces and arm reports (v0.2 aggregates only, V-para 0/58, probes 64, v0.1 60, 180 public self-tests with packs; Opus L2 rows reused)", "judge": "Claude Opus", "n": null, "seeds": null, "result": "ceiling +0.024; 0 V-para effect", "result_verified": true, "effect": "ceiling +0.024; 0 V-para effect", "ci": null, "mde": "n/a (no run)", "verdict": "measurement", "verdict_detail": "**closed by overlap, not by a measured null**: WARNSPEC v2 no-source tier is the safety line, HEALTHGAP r2 owns the classes where a memory answer is unsafe (dose / poison / co / survival), NUMGUARD r3 keeps fired numeri…", "reopen_if": "a public set where health-strict no-source turns with an empty raw list carry ≥ 0.03 pooled loss and a reviewed copy for a sourced-less answer", "closed_row": "R89", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-260", "F-1112" ], "chapter": "builds", "source": "closed registry row R89 + verified fact" }, { "id": "X-181", "code": "HEALTHDROP", "name": "Rescue of the dropped health-pack source on health-strict no-source turns", "family": "4B fast path (health route)", "question": "Rescue of the dropped health-pack source on health-strict no-source turns (HEALTHDROP, re-scoped H7; deps.healthDrop, one shared title word)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (63bfa747…)", "where_public": null }, "set": "180 public self-tests + v0.1 60 + probes 64 (gate) + v0.1/probes with health packs (stress); 0-LLM replay, detector live; 0 judge, 0 paid; criterion 63bfa747…", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "recovered 1/18", "result_verified": true, "effect": "recovered 1/18", "ci": null, "mde": "n/a (0-call gate)", "verdict": "closed", "verdict_detail": "**closed: step-0 gates (a) and (d) failed**, probe not run; the binding limits are the first-aid-only extractor and candidates outside the health packs, not the topic filter", "reopen_if": "an explanatory extractive mode for informational health questions plus a title gate stricter than one shared word, on a judged public set", "closed_row": "R90", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-1129", "F-1112" ], "chapter": "packs", "source": "closed registry row R90 + verified fact" }, { "id": "X-182", "code": "LITEREACH2-180", "name": "Wider X2b append budget on the lite tier", "family": "lite 1.7B (pack line)", "question": "Wider X2b append budget on the lite tier (LITEREACH2 → LITEREACH2-180: budget 1200, dense D 4, cap 8 vs 600/2/4 on Qwen3-1.7B Q4_K_M)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (ceabf9c0…)", "where_public": null }, "set": "180 public self-tests, 1 seed, Haiku 5.5-low N = 4 both orders (360 calls), control = LITE5-180 ON rows, same host (MacBook); 0 Opus, 0 paid; criterion ceabf9c0…", "judge": "Claude Opus, Claude Haiku (L1)", "n": null, "seeds": null, "result": "+0.005 (−0.014…+0.025)", "result_verified": true, "effect": "+0.005 (−0.014…+0.025)", "ci": "−0.014…+0.025", "mde": "0.03", "verdict": "adopted", "verdict_detail": "**closed: NO (Haiku, dev 180)**, CI excludes the +0.03 LEVER bar; the 1.7B converts the extra dense chunks at about a third of the X2b newly-present rate", "reopen_if": "a lite reader that converts present facts better (LITE6 \"thin\" bucket), or a phone prefill rate that makes +440 tokens free, plus a reach gain ≥ +40 items", "closed_row": "R91", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-1130" ], "chapter": "packs", "source": "closed registry row R91 + verified fact" }, { "id": "X-183", "code": "FOLLOWUP-r2", "name": "Article-anchored neighbour fetch for follow-ups", "family": "4B fast path (multi-turn retrieval)", "question": "Article-anchored neighbour fetch for follow-ups (FOLLOWUP r2: deps.followupAnchor, previous answer's <= 2 shown source articles, chunks scored on the bare follow-up, best 2 pinned)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (e76f0d46…)", "where_public": null }, "set": "20 agent-written public pairs + v0.1 60, 0-LLM replay, 0 judge, 0 paid; criterion e76f0d46…", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "gold kept 1 → 6/20 (gate 7)", "result_verified": true, "effect": "gold kept 1 → 6/20 (gate 7)", "ci": null, "mde": "n/a (0-call reach gate)", "verdict": "closed", "verdict_detail": "closed at reach gate (near-miss, ceiling-bound) *(review CRITIQUE-2026-10-16: below gate, no harm, ship-eligible behind a flag)*", "reopen_if": "real multi-turn traces carrying shown sources (maintainer 140-question traces or app history with titles)", "closed_row": "R92", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-556" ], "chapter": "handoff", "source": "closed registry row R92 + verified fact" }, { "id": "X-184", "code": "KIT-X2b", "name": "X2b neighbour + dense append on the claim kit", "family": "4B fast path (kit retrieval)", "question": "X2b neighbour + dense append on the claim kit (KIT-X2b: lift HEADPASS_KIT so packHeadPass a2 anchors on crypto/wikivoyage; dense half needs chunk_vecs)", "hypothesis": null, "preregistered": { "status": "see registry row", "where_public": null }, "set": "v0.2 110 kit replay traces (TK2FIX-V02 cut 7), script-only on blind, 0 LLM / 0 judge / 0 paid", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "ceiling ≈ 0.011", "result_verified": true, "effect": "ceiling ≈ 0.011", "ci": null, "mde": "0.025", "verdict": "closed", "verdict_detail": "closed at S2 (ceiling)", "reopen_if": "section-level gold on a kit set, or kit chunk_vecs rebuilt by another arm", "closed_row": "R93", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-1131", "F-1076", "F-1113" ], "chapter": "packs", "source": "closed registry row R93 + verified fact" }, { "id": "X-185", "code": "DETFP", "name": "Emergency-detector false fires on non-health v0.2 items", "family": "4B claim path", "question": "Emergency-detector false fires on non-health v0.2 items (DETFP; from WARN-DET)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (dbd13687…)", "where_public": null }, "set": "v0.2 110 (dev, scoped exception), Opus judge.v1 claim rows (2 seeds × 2 orders) + route replay cut 7 vs the PR branch; 0 calls; criterion dbd13687… before any computation", "judge": "Claude Opus", "n": null, "seeds": null, "result": "ceiling +0.0185", "result_verified": true, "effect": "ceiling +0.0185", "ci": null, "mde": "**closed: ceiling below MDE 0.02**; the PR branch (NUMGUARD r3 + HEALTHGAP r2) changes 0/7 (calculator false 7/7, no class 7/7; two also match the health lexicon). Reopen: a category/reasoning guard checked on v0.3 and V-para, never fitted on v0.2", "verdict": "closed", "verdict_detail": "arms/detfp/DETFP.md", "reopen_if": null, "closed_row": "R94", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-261" ], "chapter": "builds", "source": "closed registry row R94 + verified fact" }, { "id": "X-186", "code": "QWRITE", "name": "4B query writer for the pack search", "family": "4B fast path (pack retrieval)", "question": "4B query writer for the pack search (QWRITE: Qwen3-4B rewrites the question as a 3-8 word query for the first retrieve())", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (80695f6a…)", "where_public": null }, "set": "180 public self-tests, 0-LLM replay on cut 7 (tk2off) + 180 4B generations (seed 42, temp 0, 0.62 s/item), 0 judge, 0 paid; criterion 80695f6a…", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "top-8 69 → 67", "result_verified": true, "effect": "top-8 69 → 67", "ci": null, "mde": "**closed: step 0 FAIL**; reopen only as a new arm = two-query pool (engine + 4B query merged; union bound 87/180 top-8)", "verdict": "closed", "verdict_detail": "", "reopen_if": null, "closed_row": "R95", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-1132", "F-510" ], "chapter": "packs", "source": "closed registry row R95 + verified fact" }, { "id": "X-187", "code": "QMERGE", "name": "Two-query pool before the keep step", "family": "4B fast path (pack retrieval + keep)", "question": "Two-query pool before the keep step (QMERGE: engine query + the QWRITE 4B rewrite, both retrieve() lists merged by chunk id, unchanged keep step; QWRITE reopen)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (ea45bf53…)", "where_public": null }, "set": "180 public self-tests + v0.1 60, 0-LLM replay on pr1 8ac9eb7a + qmerge.patch e38cabc1… (detector :8094, healthgap on), QWRITE qmap + 60 new v0.1 rewrites (same writer, seed 42), 0 judge, 0 paid; crit…", "judge": "none (deterministic gate)", "n": null, "seeds": null, "result": "kept 36 → 42 (gate +10)", "result_verified": true, "effect": "kept 36 → 42 (gate +10)", "ci": null, "mde": "**closed: step 0 FAIL**; reopen only as a different merge rule (larger rewrite k / RRF interleave) under a new criterion", "verdict": "closed", "verdict_detail": "", "reopen_if": null, "closed_row": "R96", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-557", "F-510" ], "chapter": "packs", "source": "closed registry row R96 + verified fact" }, { "id": "X-188", "code": "LITE-THINK", "name": "Thinking ON for the lite 1.7B + X2b", "family": "lite 1.7B", "question": "Thinking ON for the lite 1.7B + X2b (LITE-THINK, budgets 256/512)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (64649df4…)", "where_public": null }, "set": "selftest20 (public), Haiku N=4 both orders, 80 calls, 0 Opus; criterion 64649df4… before any judged run; smoke 3/3 per budget", "judge": "Claude Opus, Claude Haiku (L1)", "n": null, "seeds": null, "result": "512: −0.041 (−0.084…+0.003); 256: −0.001", "result_verified": true, "effect": "512: −0.041 (−0.084…+0.003); 256: −0.001", "ci": "−0.084…+0.003", "mde": "**NEGATIVE / FLAT: thinking stays OFF on the lite tier**; THINK4B (4B) was NO too. Reopen: a thinking-tuned small model with a measured budget curve, or a device where TTFT is not the binding cost *(review CRITIQUE-2026-10-16: not detected at MDE ≈ 0.08 (n ≈ 20))*", "verdict": "not-detected", "verdict_detail": "arms/litethink/LITETHINK.md | relabelled per the 2026-10-16 outside critique: not detected at MDE 0.08 (n ≈ 20)", "reopen_if": null, "closed_row": "R97", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-1303", "F-667" ], "chapter": "packs", "source": "closed registry row R97 + verified fact" }, { "id": "X-189", "code": "SMALLROUTE", "name": "Small model as per-item router / classifier", "family": "4B fast path / deep routing", "question": "Small model as per-item router / classifier (SMALLROUTE S0: oracle ceilings of a lite route 1.7B+X2b vs 4B+X2b on the 180 and a deep route 4B vs deep+think 512 on v0.2; 0-call signals fic, kept_n, R2)", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen (a2fa9572…)", "where_public": null }, "set": "180 public self-tests (LITE5-180 Haiku rows) + v0.2 110 dev (cut6 vs dt512 Opus rows, keys only); 0 calls; probe criterion a2fa9572…", "judge": "Claude Opus, Claude Haiku (L1)", "n": null, "seeds": null, "result": "lite ceiling +0.018; deep signal = random at matched share", "result_verified": true, "effect": "lite ceiling +0.018; deep signal = random at matched share", "ci": null, "mde": "**NOT SEEN: no lite route (ceiling below MDE); the deep route has a real ceiling but no 0-call signal beats random at matched share**. Reopen: a trained encoder (fic labels / jev has_fact) reaching the R28 clause (≥ 70 % capture at ≤ 40 % routed) on v0.2, only with a 12 GB device plan", "verdict": "closed", "verdict_detail": "arms/smallroute/S0.md", "reopen_if": null, "closed_row": "R98", "registry_public": null, "date": "2026-10-16", "fact_ids": [ "F-558", "F-973", "F-1200" ], "chapter": "packs", "source": "closed registry row R98 + verified fact" }, { "id": "X-190", "code": "HONESTPROMPT", "name": "HONESTPROMPT: an honesty prompt for the submitted app", "family": null, "question": "Does the app maintainer's honesty prompt reduce wrong claims in long-form answers?", "hypothesis": null, "preregistered": { "status": "see arm report", "where_public": null }, "set": null, "judge": null, "n": null, "seeds": null, "result": "not available as of 2026-10-17", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": null, "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-17", "fact_ids": [], "chapter": "handoff", "source": "QUEUE (launched, no result yet)" }, { "id": "X-191", "code": "XFAM-V02", "name": "XFAM-V02: cross-family re-judge on v0.2", "family": null, "question": "Does the cross-family offset hold on v0.2?", "hypothesis": null, "preregistered": { "status": "no (post-hoc)", "where_public": null }, "set": "v0.2 (dev)", "judge": "cross-family (OpenAI gpt-5.6-sol (via the Codex CLI))", "n": null, "seeds": null, "result": "4B: raw 0.617 (0.567–0.666), floor-adjusted 0.517 (2 seeds); deep + think 512: raw 0.714 (0.666–0.759), floor-adjusted 0.639 (0.579–0.695) (1 seed); deep − 4B +0.097 vs +0.096 under Claude; Spearman 0.90–0.97; win-rate 2.7 % / 6.4 % vs 0 % under Claude", "result_verified": true, "effect": "4B: raw 0.617 (0.567–0.666), floor-adjusted 0.517 (2 seeds); deep + think 512: raw 0.714 (0.666–0.759), floor-adjusted 0.639 (0.579–0.695) (1 seed); deep − 4B +0.097 vs +0.096 under Claude; Spearman 0.90–0.97; win-rate 2.7 % / 6.4 % vs 0 % under Claude", "ci": "0.567…0.666", "mde": null, "verdict": "measurement", "verdict_detail": null, "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-17", "fact_ids": [ "F-018" ], "chapter": "measurer", "source": "canonical facts" }, { "id": "X-192", "code": "PACKS-L2", "name": "PACKS-L2", "family": "judge-protocol / packs", "question": "Do the pack line's Haiku L1 paired cells (S − OFF, P − S, X − P on the 180 self-tests; wave-3 S − OFF) hold under Opus judge.v1 L2 on the same traces?", "hypothesis": null, "preregistered": { "status": "yes, sha256-frozen", "where_public": null }, "set": "180 self-tests; wave-3 self-tests", "judge": "Claude Opus", "n": "180 (wave 3 n 40 separate); ≤ 1,300 Opus calls", "seeds": "1 (judge seed 7)", "result": "Opus +0.042 (+0.024…+0.062); Haiku +0.032 (+0.016…+0.050)", "result_verified": true, "effect": "Opus +0.042 (+0.024…+0.062); Haiku +0.032 (+0.016…+0.050)", "ci": "+0.024…+0.062", "mde": null, "verdict": "closed", "verdict_detail": "closed: S − OFF CONFIRMED, P − S COST CONFIRMED, X − P OPTIONAL (R49 unchanged)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": null, "fact_ids": [ "F-025", "F-026" ], "chapter": "packs", "source": "arm-index v2 (ledger audit 2026-10-13) + verified facts" }, { "id": "X-193", "code": "PROBES", "name": "PROBES", "family": "tooling", "question": "Data-only directory: dev and held-out probe sets used by the cut3-dev AR (arithmetic), SC (water-safety card) and VG (verdict guard) L0 gates; no criterion or result inside, verdicts live in arms/CUT3DEV.md and arms/{ar,sc}-L0-criterion.md.", "hypothesis": null, "preregistered": { "status": "yes, frozen file without a sha256 sidecar", "where_public": null }, "set": "cut3-dev probe sets (arithmetic, water-safety, vg-questions)", "judge": "none (deterministic gate)", "n": "arithmetic.probes 45, arithmetic.heldout 50, vg-questions.probes 30, water-safety.probes 40, water-safety.heldout 50 (l…", "seeds": "n/a", "result": "n/a (consumers: AR A6 held-out recall 28/32, 0 wrong numbers, 0/18 false fires PASS; SC S5 held-out recall 18/25 = 72 % FAIL, 0/25 false fires)", "result_verified": false, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "tooling (data-only probe sets for the cut3-dev AR/SC/VG L0 gates; results in arms/CUT3DEV.md)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": null, "fact_ids": [], "chapter": "builds", "source": "arm-index v2 (ledger audit 2026-10-13); result not re-verified for this publication" }, { "id": "X-198", "code": "HONESTPROMPT", "name": "HONESTPROMPT", "family": "packs-delivery", "question": "Do honesty clauses in the Detailed prompt cut invented facts without refusals?", "hypothesis": null, "preregistered": { "status": "yes (5c73a30f + addenda)", "where_public": null }, "set": "set A 15 Detailed turns; set B 20 v0.1 no-retrieval ids; v0.1 57; 49 capability turns (dev)", "judge": "Claude Opus sentence audit (wrong claims / claims); Claude Haiku ratio", "n": null, "seeds": null, "result": "wrong D0 0.231 → D1 0.062 (set A); pooled A+B 0.216→0.098; refusals D1 2/49, S1 7/49; code blocks 8/21→0/21", "result_verified": true, "effect": null, "ci": "set A −0.169 (−0.281..−0.073); pooled −0.118 (−0.180..−0.057); v0.1 ratio −0.027 (−0.060..+0.006)", "mde": null, "verdict": "closed", "verdict_detail": "D1 GO (code caveat); S1 NO; D2 partial", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-18", "fact_ids": [ "F-1866" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-199", "code": "HONESTPROMPT-D4", "name": "HONESTPROMPT-D4", "family": "prompting-style", "question": "Is D4 (D1 + capability sentence, no \"examples\") an upgrade?", "hypothesis": null, "preregistered": { "status": "yes (addendum 3)", "where_public": null }, "set": "set A 15; v0.1 57 (dev)", "judge": "Claude Opus audit; Claude Haiku ratio", "n": null, "seeds": null, "result": "wrong 0.065 at 11.2 claims/answer; v0.1 D4 0.427 vs D1 0.425; code 2/21", "result_verified": true, "effect": null, "ci": "D4−D0 wrong −0.166 (−0.270..−0.092); v0.1 D4−D1 +0.003 (−0.024..+0.028)", "mde": null, "verdict": "adopted", "verdict_detail": "D4 recommended Detailed prompt", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-18", "fact_ids": [ "F-1867" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-200", "code": "CHATFAIL-S0", "name": "CHATFAIL-S0", "family": null, "question": "Do intent (L2) and follow-up anchor (L3) rules fix code refusals and topic loss?", "hypothesis": null, "preregistered": { "status": "no (probe sha 30317276, 0 calls)", "where_public": null }, "set": "CHAT-FU-v1 30 chats / 159 turns (dev)", "judge": "routing replay", "n": null, "seeds": null, "result": "intent 74.8→85.5 %; code asks 0→81 % (17/21); anchor turn 3+ 20→77 %", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "PARTIAL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-18", "fact_ids": [ "F-1868" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-201", "code": "CHATFAIL-S1", "name": "CHATFAIL-S1", "family": "models", "question": "Do the routing fixes raise fulfilled capability turns on 4B and 1.5B?", "hypothesis": null, "preregistered": { "status": "detectors hashed before generation", "where_public": null }, "set": "CHAT-FU-v1 49 capability turns; v0.1 57 non-regression (dev)", "judge": "strict detector + Claude Opus audit; Claude Haiku v0.1", "n": null, "seeds": null, "result": "4B fulfilled 0.633→0.755; 1.5B +0.020; refusals 0.061→0; Opus: real gain ~+0.10, 1/20 code runnable; 1.5B adds 5 wrong first-aid steps", "result_verified": true, "effect": null, "ci": "4B +0.122 (0.000..+0.256); v0.1 −0.014 (−0.029..+0.001)", "mde": null, "verdict": "open", "verdict_detail": "4B PARTIAL; 1.5B NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-18", "fact_ids": [ "F-1869" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-204", "code": "LONGDIGIT", "name": "LONGDIGIT", "family": null, "question": "Are engine number regexes safe on long digit runs?", "hypothesis": null, "preregistered": { "status": "n/a (bug repro + test)", "where_public": null }, "set": "synthetic strings", "judge": "timing", "n": null, "seeds": null, "result": "8,000 digits 714 ms; fixed with digit lookbehinds; all exports <50 ms at 20k", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "patch shipped in PR 85", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-18", "fact_ids": [ "F-1872" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-205", "code": "COCAR", "name": "COCAR", "family": null, "question": "Does the CO card catch fuel heaters in vehicles?", "hypothesis": null, "preregistered": { "status": "n/a", "where_public": null }, "set": "625 public questions + tests", "judge": "false-positive count", "n": null, "seeds": null, "result": "0/625 public FP change; tests 12/12", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "patch shipped in PR 85", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-18", "fact_ids": [ "F-1873" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-206", "code": "FACTS-R7 / FACTSAUDIT / R8 / R9 / R11", "name": "FACTS-R7 / FACTSAUDIT / R8 / R9 / R11", "family": "packs-delivery", "question": "Is the offline facts table correct?", "hypothesis": null, "preregistered": { "status": "yes (FACTSAUDIT 7b2042dc", "where_public": null }, "set": "197 countries vs a public factbook; 625-675 public questions", "judge": "deterministic audit", "n": null, "seeds": null, "result": "r7: 207 bare-QID labels dropped; audit 33 flags, 5 table-wrong; r8 0 table-wrong + emergency numbers for 195 countries; r9 adds driving side, calling code, mains Hz, 0 wrong, FP 0/675; r11 root-cause builder fix", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "shipped as handoffs", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-18", "fact_ids": [ "F-1874" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-207", "code": "PROMPT3M", "name": "PROMPT3M", "family": "packs-delivery", "question": "Same honesty prompts on 4B, LFM2.5, 1.7B (no source)", "hypothesis": null, "preregistered": { "status": "yes (50da1b4b)", "where_public": null }, "set": "35 items + capability gate (dev)", "judge": null, "n": null, "seeds": null, "result": "stopped: chain overlapped PACKREAD; 1.7B main set complete, gates not run", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "incomplete", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-19", "fact_ids": [ "F-1875" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-208", "code": "PACKREAD", "name": "PACKREAD", "family": "packs-delivery", "question": "With the matching pack and the engine's own contexts, which model + prompt reads best?", "hypothesis": null, "preregistered": { "status": "yes (4649513d)", "where_public": null }, "set": "packread-v1 54 pack-matched self-tests (dev, pack-written)", "judge": "Claude Haiku vs pack refs", "n": null, "seeds": null, "result": "engine Succinct 0.610; 4B D0/D4/D5 0.640/0.650/0.657; LFM 0.539/0.554/0.559; 1.7B 0.485/0.529/0.522", "result_verified": true, "effect": null, "ci": "4B-D4 vs engine +0.041 (+0.009..+0.076); LFM-D4 − 4B-D4 −0.096 (−0.145..−0.052); 1.7B-D0 − engine −0.125 (−0.165..−0.087)", "mde": null, "verdict": "adopted", "verdict_detail": "4B + D4 best; honest prompts never hurt", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-19", "fact_ids": [ "F-1876" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-209", "code": "SAFEPACK-E", "name": "SAFEPACK-E", "family": "packs-delivery", "question": "Does the engine answer emergencies safely with the emergency pack installed?", "hypothesis": null, "preregistered": { "status": "rule SAFE-FAIL ≤5", "where_public": null }, "set": "safepack 30 agent-written emergency items (dev)", "judge": "blind Claude Opus safety audit", "n": null, "seeds": null, "result": "unsafe 0, declined 20, missing-critical 26", "result_verified": true, "effect": null, "ci": "SAFE-FAIL 26/30 vs ≤5", "mde": null, "verdict": "closed", "verdict_detail": "FAIL (safe, not helpful)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-19", "fact_ids": [ "F-1877" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-210", "code": "SAFEROUTE", "name": "SAFEROUTE", "family": "packs-delivery", "question": "Does searching \" what to do\" bring the official page into sources?", "hypothesis": null, "preregistered": { "status": "yes (841da944)", "where_public": null }, "set": "safepack 30; packread 54; v0.1 60 (dev)", "judge": "replay; then blind Claude Opus", "n": null, "seeds": null, "result": "gold in sources 1→21/30; empty retrieval 10→2; refusals 18→2; but unsafe extracts 0→3 (corrected to 2)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "retrieval PASS; safety rule FAIL; held from PR", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-19", "fact_ids": [ "F-1878" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-211", "code": "SAFEPACK-ORACLE", "name": "SAFEPACK-ORACLE", "family": "packs-delivery", "question": "With the right page handed over, which model answers emergencies completely and safely?", "hypothesis": null, "preregistered": { "status": "addendum 971f94ab", "where_public": null }, "set": "safepack 30 × 6 arms (dev)", "judge": "blind Claude Opus", "n": null, "seeds": null, "result": "complete-and-safe 4B-D4 23, 4B-D0 18, LFM 11-14, 1.7B 9-13; unsafe 4B-D4 3, LFM 11-13", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "no arm meets rule; LFM least safe", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-19", "fact_ids": [ "F-1879" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24" }, { "id": "X-212", "code": "WATERCARD", "name": "WATERCARD", "family": "safety", "question": "Revive the drinking-water safety card with a v3 trigger", "hypothesis": null, "preregistered": { "status": "n/a (tests)", "where_public": null }, "set": "640 public + earlier probe sets", "judge": "fire counts", "n": null, "seeds": null, "result": "public fires 2/640; old recall probes 22/22", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "patch 7 built", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-19", "fact_ids": [ "F-1880" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-213", "code": "SMALLCLS", "name": "SMALLCLS", "family": "models", "question": "Can a 1.7B classifier replace the hand hazard list?", "hypothesis": null, "preregistered": { "status": "yes (132fa1f0)", "where_public": null }, "set": "85 fresh hazard texts + section choice (dev)", "judge": "deterministic", "n": null, "seeds": null, "result": "recall 38/45 vs hand 7/45, 0 false alarms; corrected right-hazard 14/45 (4B 32/45); section choice weak", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "not adopted", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-19", "fact_ids": [ "F-1881" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-214", "code": "RELGATE", "name": "RELGATE", "family": "ruler", "question": "Can the 4B judge relevance / sufficiency / source pick in-harness?", "hypothesis": null, "preregistered": { "status": "yes (10b462f8)", "where_public": null }, "set": "150 contexts / 642 chunks, public self-test ground truth, held-out half B", "judge": "AUC / accuracy", "n": null, "seeds": null, "result": "source pick 72 % (31/43) vs 37 % baseline; per-chunk AUC 0.83 FAIL; sufficiency AUC 0.72 FAIL; 1.7B useless", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "G3 PASS only", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-19", "fact_ids": [ "F-1882" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-215", "code": "PACK-NIST", "name": "PACK-NIST", "family": "packs-delivery", "question": "Does a public-domain NIST crypto pack help?", "hypothesis": null, "preregistered": { "status": "criterion in arm", "where_public": null }, "set": "20 public self-tests (pack-written)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "0.430 vs 0.422; pack kept 3/20", "result_verified": true, "effect": null, "ci": "+0.008 (−0.012..+0.030)", "mde": null, "verdict": "measurement", "verdict_detail": "user-selected only", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-19", "fact_ids": [ "F-1883" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-216", "code": "CHAINWHEN", "name": "CHAINWHEN", "family": "packs-delivery", "question": "Does chaining the 4B's source pick before the answer help?", "hypothesis": null, "preregistered": { "status": "yes (adba95cb)", "where_public": null }, "set": "86 triggered of 150 pack contexts (dev)", "judge": "Claude Haiku, 472 calls", "n": null, "seeds": null, "result": "+0.019 on triggered; overall +0.011 (−0.003..+0.025)", "result_verified": true, "effect": null, "ci": "+0.019 (−0.006..+0.044)", "mde": null, "verdict": "closed", "verdict_detail": "NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1884" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-217", "code": "VEXACT-ACCEPT", "name": "VEXACT-ACCEPT", "family": null, "question": "Baseline for a venue acceptance harness", "hypothesis": null, "preregistered": { "status": "yes (e762c3b0)", "where_public": null }, "set": "V-exact 11 (dev)", "judge": "Claude Haiku, 2 seeds", "n": null, "seeds": null, "result": "0.475; venues 6/6 both seeds", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "baseline set", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1885" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-218", "code": "PACK-VISA", "name": "PACK-VISA", "family": "packs-delivery", "question": "Does a visa-articles pack help?", "hypothesis": null, "preregistered": { "status": "criterion in arm", "where_public": null }, "set": "20 self-tests (pack-written)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "0.366 vs 0.359; kept 1/20 (destination pages outrank)", "result_verified": true, "effect": null, "ci": "+0.007 (−0.024..+0.041)", "mde": null, "verdict": "tooling", "verdict_detail": "built, inert; default-eligible by absence", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1886" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-220", "code": "LITE17", "name": "LITE17", "family": "packs-delivery", "question": "Do thinking@256, quote-then-answer or SE question-chunk drop close the 1.7B reading gap?", "hypothesis": null, "preregistered": { "status": "criterion in arm", "where_public": null }, "set": "150 pack items (dev, pack-written)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "1.7B 0.50 vs 4B 0.61 unchanged", "result_verified": true, "effect": null, "ci": "T −0.002 (−0.023..+0.019); Q +0.002 (−0.018..+0.024); S +0.001 (−0.028..+0.032)", "mde": null, "verdict": "closed", "verdict_detail": "NO (lite reading levers closed)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1888" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-221", "code": "PLUGBATCH v1", "name": "PLUGBATCH v1", "family": null, "question": "Build an adversarial plugin batch", "hypothesis": null, "preregistered": { "status": "n/a", "where_public": null }, "set": "110 public rows", "judge": "deterministic", "n": null, "seeds": null, "result": "TS plugin functions alone break 23 rows", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "built", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1889" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-222", "code": "LITEDECLINE", "name": "LITEDECLINE", "family": "ruler", "question": "Do the 1.7B's compact decline rules cost utility?", "hypothesis": null, "preregistered": { "status": "yes", "where_public": null }, "set": "v0.1 57 + 180 self-tests (dev)", "judge": "Claude Haiku, 678 calls", "n": null, "seeds": null, "result": "declines 94/237; 3 named rules off: Q +0.005, U0 +0.014; all 5 off (report-only): U0 +0.170 (+0.141..+0.199), CE rate +0.025 (−0.013..+0.063)", "result_verified": true, "effect": null, "ci": "U0 +0.014 (+0.005..+0.024)", "mde": null, "verdict": "open", "verdict_detail": "PARTIAL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1890" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-223", "code": "LITELAT", "name": "LITELAT", "family": "packs-delivery", "question": "Where does 1.7B time-to-first-token go on a Pixel 8a?", "hypothesis": null, "preregistered": { "status": "yes (22c987fc)", "where_public": null }, "set": "litelat-9 public dev, 7 phone runs", "judge": "timing", "n": null, "seeds": null, "result": "TTFT p50 9.4 s = retrieval 5.3 s (435 serial SQL calls) + prefill 3.3 s; no runtime override passes", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "diagnosis", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1891" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-225", "code": "CMP30", "name": "CMP30", "family": "ruler", "question": "1.7B vs 1.5B on Pixel 8a with product decline rules", "hypothesis": null, "preregistered": { "status": "no (measurement)", "where_public": null }, "set": "30 public v0.1, 26 judged (dev)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "Q 0.301 vs 0.334; 1.7B declines 17 vs 13 of 26; 1.5B faster", "result_verified": true, "effect": null, "ci": "Q +0.033 (−0.017..+0.078); answered correctness +0.070 (+0.014..+0.129, n 8)", "mde": null, "verdict": "measurement", "verdict_detail": "measurement", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1893" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-226", "code": "BATCHRET", "name": "BATCHRET", "family": "packs-delivery", "question": "Batch the pack SQL lookups without changing retrieval", "hypothesis": null, "preregistered": { "status": "yes (b177323c)", "where_public": null }, "set": "390 public dev items", "judge": "identity + timing", "n": null, "seeds": null, "result": "SQL 937→48 per retrieve (−94.9 %); identical 390/390", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "PASS", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1894" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-227", "code": "CMP30OFF", "name": "CMP30OFF", "family": "ruler", "question": "With declines released behind a warning, 1.7B vs 1.5B on phone", "hypothesis": null, "preregistered": { "status": "no (measurement)", "where_public": null }, "set": "30 public v0.1, 26 judged", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "U0 1.7B 0.394 vs 1.5B 0.349; CE 8 vs 10; half the warned answers are still CEs", "result_verified": true, "effect": null, "ci": "Q +0.045 (−0.009..+0.100)", "mde": null, "verdict": "measurement", "verdict_detail": "measurement", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1895" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-228", "code": "BATCHRET-PHONE", "name": "BATCHRET-PHONE", "family": "packs-delivery", "question": "Measured phone gain of batched lookups", "hypothesis": null, "preregistered": { "status": "same as BATCHRET", "where_public": null }, "set": "litelat-9, Pixel 8a, A/B/A/B", "judge": "timing", "n": null, "seeds": null, "result": "1.7B TTFT 9.12→5.55 s; retrieval 5.04→1.45 s; identical 9/9", "result_verified": true, "effect": null, "ci": "TTFT −3.49 s (−39 %)", "mde": null, "verdict": "adopted", "verdict_detail": "PASS", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1896" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-229", "code": "LP-SCREENS", "name": "LP-SCREENS", "family": "packs-delivery", "question": "Set up K0/K1 problem screens with sentinels on one host", "hypothesis": null, "preregistered": { "status": "setup", "where_public": null }, "set": "27 problems + 10 sentinels per screen (dev)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "host parity text 11-29 %; retrieval identical; sentinel rule raised to 1.0 correctness drop", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "harness", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1897" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-230", "code": "B-HONEST", "name": "B-HONEST", "family": "prompting-style", "question": "D4 honest clauses on the 1.7B", "hypothesis": null, "preregistered": { "status": "yes (3ac3efca)", "where_public": null }, "set": "K0 screen 37 (dev)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "fixed 4 of 23 CE (need 6)", "result_verified": true, "effect": null, "ci": "U +0.012 (−0.015..+0.043)", "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1898" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-231", "code": "SHORTCTX", "name": "SHORTCTX", "family": "packs-delivery", "question": "≤2 chunks of context for the 1.7B", "hypothesis": null, "preregistered": { "status": "yes (865cb4f5)", "where_public": null }, "set": "K1 screen 37", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "fixed 4 of 21", "result_verified": true, "effect": null, "ci": "U −0.018 (−0.047..+0.010)", "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1899" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-232", "code": "QUOTEFIRST", "name": "QUOTEFIRST", "family": null, "question": "Open with the source sentence", "hypothesis": null, "preregistered": { "status": "yes (604dec4d)", "where_public": null }, "set": "K1 screen 37", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "fixed 4 of 21; model paraphrases", "result_verified": true, "effect": null, "ci": "U −0.011 (−0.058..+0.028)", "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1900" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-233", "code": "CE-TAXONOMY", "name": "CE-TAXONOMY", "family": "ruler", "question": "Why does the 1.7B make confident errors?", "hypothesis": null, "preregistered": { "status": "no (labels, leads)", "where_public": null }, "set": "27 K0 problem items", "judge": "Claude Haiku labels", "n": null, "seeds": null, "result": "20/27 had no relevant source; 6 reading failures; fix = new pack content 15", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "lead", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1901" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-234", "code": "GREEDY", "name": "GREEDY", "family": "models", "question": "Greedy decoding on the 1.7B", "hypothesis": null, "preregistered": { "status": "yes (7e17dd98)", "where_public": null }, "set": "K0 screen 37", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "fixed 4 of 23", "result_verified": true, "effect": null, "ci": "CE −0.108 (−0.216..−0.027); U +0.004", "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1902" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-235", "code": "LOGPROB", "name": "LOGPROB", "family": null, "question": "Does token confidence flag confident errors?", "hypothesis": null, "preregistered": { "status": "yes (c80540f9)", "where_public": null }, "set": "MAIN 97 regenerated with logprobs", "judge": "AUC vs Claude Haiku labels", "n": null, "seeds": null, "result": "best AUC 0.530 (0.400-0.657)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1903" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-236", "code": "K1-COVERAGE", "name": "K1-COVERAGE", "family": "packs-delivery", "question": "Are the 1.7B's no-source errors retrieval misses?", "hypothesis": null, "preregistered": { "status": "no (0 calls)", "where_public": null }, "set": "20 no-source CEs", "judge": "read-only FTS", "n": null, "seeds": null, "result": "key fact in neither new pack 10/20; 4 blocked before retrieval", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "content first", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1904" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-237", "code": "PHIFIRST", "name": "PHIFIRST", "family": null, "question": "Does first-token entropy flag CEs?", "hypothesis": null, "preregistered": { "status": "yes (3ff3394f)", "where_public": null }, "set": "97 LOGPROB items", "judge": "AUC", "n": null, "seeds": null, "result": "0.488 (0.368-0.601); inverted on generated items", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1905" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-238", "code": "VENDORSAMP", "name": "VENDORSAMP", "family": "models", "question": "Qwen's recommended non-thinking sampling on the 1.7B", "hypothesis": null, "preregistered": { "status": "yes (666dc802)", "where_public": null }, "set": "K0 screen 37", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "fixed 1 of 23; 22/37 answers identical", "result_verified": true, "effect": null, "ci": "CE −0.027 (−0.081..0.000)", "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1906" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-239", "code": "OCCRAG", "name": "OCCRAG", "family": "models", "question": "OCC-RAG-1.7B (faithful-reading fine-tune) as hybrid reader", "hypothesis": null, "preregistered": { "status": "yes (63754117)", "where_public": null }, "set": "K0, K1 screens", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "answers 10 of 40 eligible, refuses 22, breaks format 8; 2x slower", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1907" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-240", "code": "MODEL-USAGE audit", "name": "MODEL-USAGE audit", "family": "models", "question": "Is the measurement pipeline sound?", "hypothesis": null, "preregistered": { "status": "n/a (7 finders + 2 refuters)", "where_public": null }, "set": null, "judge": "review", "n": null, "seeds": null, "result": "33 findings, 6 P0 (stale CE metric; refusals labelled CE; unchecked X2b padding; unreachable screen bar; Qwen3.5-2B unscreened)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "follow-ups launched", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1908" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-241", "code": "REREAD", "name": "REREAD", "family": null, "question": "Re-read screens in refuse-by-default mode", "hypothesis": null, "preregistered": { "status": "no (0 calls)", "where_public": null }, "set": "5 screens", "judge": "re-count", "n": null, "seeds": null, "result": "only 4 visible CE per screen; quote-first is a regression (4→8)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "frozen verdicts unchanged", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1909" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-242", "code": "GATE17", "name": "GATE17", "family": "models", "question": "Groundedness LoRA as a CE detector", "hypothesis": null, "preregistered": { "status": "yes (84cb4fa0)", "where_public": null }, "set": "57 source-present answers", "judge": "AUC vs correctness", "n": null, "seeds": null, "result": "0.496 (0.340-0.648)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO (measured the wrong target)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1910" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-243", "code": "LITEPACKS", "name": "LITEPACKS", "family": "packs-delivery", "question": "Do K1-K4 pack sets change the 1.7B?", "hypothesis": null, "preregistered": { "status": "yes (frozen + secondary)", "where_public": null }, "set": "MAIN 97; 80 pack-matched self-tests; 100 SE off-pack (dev)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "MAIN no change; self-tests K3 +0.069, K2 +0.045, K1 +0.036; no crowding flag", "result_verified": true, "effect": null, "ci": "K3 +0.069 (+0.038..+0.099)", "mde": null, "verdict": "closed", "verdict_detail": "MAIN NO CHANGE", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1911" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-244", "code": "GATE17-SUPPORT", "name": "GATE17-SUPPORT", "family": "models", "question": "Does the LoRA separate supported from unsupported answers?", "hypothesis": null, "preregistered": { "status": "yes (98750b05)", "where_public": null }, "set": "57 source-present answers; Haiku support labels (κ 0.79)", "judge": "AUC", "n": null, "seeds": null, "result": "0.965 (0.913-1.000); 0/9 false refusals; 84 % unsupported", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1912" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-245", "code": "RETFLOOR", "name": "RETFLOOR", "family": "ruler", "question": "Remove off-topic appended chunks (V1orC rule)", "hypothesis": null, "preregistered": { "status": "criterion in arm", "where_public": null }, "set": "MAIN 97 + 180 replays (0 LLM)", "judge": "counts", "n": null, "seeds": null, "result": "V1orC removes all 82 off-topic appends, keeps fact reach 98 vs 97; 8.2 pts of K1 coverage were padding", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "rule found; screened later (NO-GO)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1913" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-246", "code": "JUDGE-V2", "name": "JUDGE-V2", "family": "ruler", "question": "Binary key-point checklist judge vs judge.v1", "hypothesis": null, "preregistered": { "status": "yes (b7a1b12f)", "where_public": null }, "set": "3×3 repeats on MAIN 97 control answers", "judge": "test-retest", "n": null, "seeds": null, "result": "CE retest 0.966 vs v1 0.924 (needed 0.974); score SD 0.80x v1 (needed 0.60x)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-GO; judge.v1 stays", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1914" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-247", "code": "CE-OPUS", "name": "CE-OPUS", "family": "ruler", "question": "Do Haiku CE labels hold under Opus?", "hypothesis": null, "preregistered": { "status": "yes (98a826d9)", "where_public": null }, "set": "47 answers", "judge": "Claude Haiku vs Claude Opus", "n": null, "seeds": null, "result": "CE agree 39/47 (κ 0.63); CE-unwarned 45/47 (κ 0.87); Haiku over-calls CE", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "default-mode metric judge-robust", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1915" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-248", "code": "PACK-HOLES", "name": "PACK-HOLES", "family": "packs-delivery", "question": "Do stripped wiki templates leave fact holes?", "hypothesis": null, "preregistered": { "status": "no (0 LLM scan)", "where_public": null }, "set": "pack files", "judge": "scan", "n": null, "seeds": null, "result": "holes simplewiki 1.29 %, wikivoyage 1.36 %; 3.2 % of K1 v0.1 kept chunks", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "converter bug found", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1916" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-249", "code": "REFUSE-UNSUPPORTED", "name": "REFUSE-UNSUPPORTED", "family": "ruler", "question": "1.7B + groundedness gate refusing unsupported answers", "hypothesis": null, "preregistered": { "status": "yes (af16f29c)", "where_public": null }, "set": "177 = MAIN 97 + 80 pack-matched self-tests", "judge": "Claude Haiku; Claude Opus re-judge (CE-OPUS-RU)", "n": null, "seeds": null, "result": "CE-unwarned 10→4 of 177 (Opus 7→3); false refusals 0/15 (Opus 0/14); coverage 0.435→0.328; projected +9.3 s per answer on Pixel 8a", "result_verified": true, "effect": null, "ci": "U on answered +0.030 (+0.005..+0.057); Opus CE diff CI −0.051..−0.006", "mde": null, "verdict": "closed", "verdict_detail": "PASS (Haiku and Opus); latency blocks shipping", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1917" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-250", "code": "QUANT17", "name": "QUANT17", "family": "prompting-style", "question": "Is Q4_K_M close enough to BF16 for the 1.7B?", "hypothesis": null, "preregistered": { "status": "yes", "where_public": null }, "set": "30 prompts, 1,370 tokens", "judge": "KLD / top-token", "n": null, "seeds": null, "result": "Q4 KLD 0.148, top-token 91.2 %; Q8 0.007 / 97.2 %", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "Q4 fails KLD rule → Q8 screen", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1918" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-251", "code": "NULL-CAL / GATE v2", "name": "NULL-CAL / GATE v2", "family": null, "question": "How much do null arms \"fix\" on a screen?", "hypothesis": null, "preregistered": { "status": "yes (f8239d8b, d1d4b6e4)", "where_public": null }, "set": "new K0 30 / K1 21 problems + 20 sentinels", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "nulls fix 23-31 %; net 6-8; bar = max null net + 2 (K0 10, K1 9)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "new screen rule", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1919" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-252", "code": "RETV1ORC", "name": "RETV1ORC", "family": null, "question": "Retrieval hygiene (V1orC) as a quality lever", "hypothesis": null, "preregistered": { "status": "yes (0f2a7016)", "where_public": null }, "set": "GATE v2 K1 screen 41", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "net 2 (bar 9)", "result_verified": true, "effect": null, "ci": "U on covered −0.026 (−0.064..+0.004)", "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1920" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-253", "code": "GATELAT", "name": "GATELAT", "family": "packs-delivery", "question": "A cheaper gate layout that keeps quality", "hypothesis": null, "preregistered": { "status": "yes (94e86e40)", "where_public": null }, "set": "124 gated answers", "judge": "AUC", "n": null, "seeds": null, "result": "context-first + KV reuse 1.8 s but AUC 0.934→0.660", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1921" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-254", "code": "MINISTRAL-3-3B", "name": "MINISTRAL-3-3B", "family": "models", "question": "Ministral-3-3B as the 8 GB model", "hypothesis": null, "preregistered": { "status": "yes (fa1c4a9b)", "where_public": null }, "set": "GATE v2 K0/K1", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "nets clear bars but false refusals 3-9/20 from dropped [n] citations", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-GO (citation format)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1922" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-255", "code": "SHORTCTX2", "name": "SHORTCTX2", "family": null, "question": "Short context under GATE v2", "hypothesis": null, "preregistered": { "status": "yes (97b055a2)", "where_public": null }, "set": "K1 screen 41", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "net 8 (bar 9); false refusals 6/20", "result_verified": true, "effect": null, "ci": "U cov −0.132 (−0.222..−0.053)", "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1923" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-256", "code": "GEMMA4-E2B", "name": "GEMMA4-E2B", "family": "models", "question": "Gemma 4 E2B as the 8 GB model", "hypothesis": null, "preregistered": { "status": "yes (690eeead)", "where_public": null }, "set": "4 GATE v2 screens; MAIN 97 stage 2", "judge": "Claude Haiku; later Claude Opus", "n": null, "seeds": null, "result": "screens GO (net 16-21); stage 2 U 0.4055→0.4336", "result_verified": true, "effect": null, "ci": "dU +0.028 (+0.006..+0.052), needed +0.03", "mde": null, "verdict": "closed", "verdict_detail": "stage-2 FAIL by 0.002 (Haiku)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-20", "fact_ids": [ "F-1924" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-257", "code": "GEMMA4-CONFIRM", "name": "GEMMA4-CONFIRM", "family": "models", "question": "Seed 43 and K1 for Gemma", "hypothesis": null, "preregistered": { "status": "yes (399cfe86, measurement)", "where_public": null }, "set": "MAIN 97", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "seed 43 +0.030 (+0.007..+0.052); pooled +0.029 (+0.009..+0.048); K1 +0.017 (−0.008..+0.040); gain = fewer refusals", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "no verdict change", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1925" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-258", "code": "GEMMA-OPUS-MAIN", "name": "GEMMA-OPUS-MAIN", "family": "models", "question": "Gemma vs 1.7B under Opus", "hypothesis": null, "preregistered": { "status": "yes (8bf4a4ae)", "where_public": null }, "set": "MAIN 97", "judge": "Claude Opus, 362 calls", "n": null, "seeds": null, "result": "s42 U 0.435→0.466; s43 +0.041; pooled +0.036 (+0.011..+0.062)", "result_verified": true, "effect": null, "ci": "s42 +0.031 (+0.005..+0.058)", "mde": null, "verdict": "adopted", "verdict_detail": "GO by 0.001 (Opus)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1926" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-259", "code": "Q8SCREEN", "name": "Q8SCREEN", "family": "models", "question": "Q8_0 instead of Q4_K_M", "hypothesis": null, "preregistered": { "status": "yes (22e97896)", "where_public": null }, "set": "GATE v2 K0/K1", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "net 9 (bar 10), 7 (bar 9); decode x0.66; +0.73 GB", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-GO; Q4 stays", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1927" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-260", "code": "QWEN35-2B", "name": "QWEN35-2B", "family": "models", "question": "Qwen3.5-2B as the 8 GB model", "hypothesis": null, "preregistered": { "status": "yes (219ef0e3)", "where_public": null }, "set": "4 screens", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "drops [n] citations; false refusals up to 8/20; answers ~2x longer", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1928" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-261", "code": "OCCDECL", "name": "OCCDECL", "family": "models", "question": "OCC-RAG 1.7B/0.6B with abstain-as-refusal", "hypothesis": null, "preregistered": { "status": "yes (01b04fc8)", "where_public": null }, "set": "4 screens", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "fixes mostly by refusing; K1 7-9 false refusals", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-GO; OCC line closed", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1929" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-262", "code": "FOTAPOL-LORA", "name": "FOTAPOL-LORA", "family": "models", "question": "A public RAG LoRA for Qwen3-1.7B", "hypothesis": null, "preregistered": { "status": "yes (d00e5fa9)", "where_public": null }, "set": "2 screens", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "net 6 (bar 10), 9 (bar 9) with FR 4/20", "result_verified": true, "effect": null, "ci": "U cov −0.084", "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1930" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-263", "code": "CITEFIX", "name": "CITEFIX", "family": null, "question": "Is the citation-drop failure a format issue?", "hypothesis": null, "preregistered": { "status": "yes (93c85722)", "where_public": null }, "set": "910 candidate answers", "judge": "0 Claude Haiku", "n": null, "seeds": null, "result": "engine's 50 % key-word support check strips 161 paraphrased citations; format fix flips no gate", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "open lever: paraphrase-tolerant check", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1931" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-264", "code": "JAN-V1-EDGE", "name": "JAN-V1-EDGE", "family": "models", "question": "Jan-v1-edge (1.7B SFT+RL) thinking off", "hypothesis": null, "preregistered": { "status": "yes (3a1740ea)", "where_public": null }, "set": "4 screens", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "K0 net 8 and 7 (bar 10); U_cov fails everywhere", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1932" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-265", "code": "MIN3CITE", "name": "MIN3CITE", "family": "prompting-style", "question": "Ministral-3-3B + citation prompt v2", "hypothesis": null, "preregistered": { "status": "yes (a4d2d9a3", "where_public": null }, "set": "screens; MAIN 97 K0", "judge": "Claude Haiku; Claude Opus revisit", "n": null, "seeds": null, "result": "stage 2 Haiku +0.062 (+0.037..+0.089); Opus +0.047 (+0.021..+0.075); ~3x phone time; −0.037 vs 4B (Opus, 61)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "Haiku PASS; closed as dominated", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1933" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-266", "code": "GEMMA4-PHONE", "name": "GEMMA4-PHONE", "family": "models", "question": "Gemma on Pixel 8a: speed, memory, quality", "hypothesis": null, "preregistered": { "status": "judge criterion 54cf6c77", "where_public": null }, "set": "73 phone items; 57 v0.1 parity ids", "judge": "timing; Claude Haiku", "n": null, "seeds": null, "result": "73/73, p50 11.8 s, decode 7.0 tok/s, PSS 4.52 GB; phone Gemma 0.452 vs phone 4B 0.535; phone = desktop Gemma (+0.001)", "result_verified": true, "effect": null, "ci": "Gemma − 4B −0.083 (−0.123..−0.047)", "mde": null, "verdict": "measurement", "verdict_detail": "measurement", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1934" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-267", "code": "SMOLLM3-3B", "name": "SMOLLM3-3B", "family": "models", "question": "SmolLM3-3B", "hypothesis": null, "preregistered": { "status": "yes (a9f5143d)", "where_public": null }, "set": "4 screens", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "false refusals 3-7/20, all dropped citations", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1935" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-268", "code": "GRANITE4-1B", "name": "GRANITE4-1B", "family": "models", "question": "granite-4.0-1b", "hypothesis": null, "preregistered": { "status": "yes (ed62c89a)", "where_public": null }, "set": "screens; MAIN 97", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "stage 2 U 0.4055→0.4058", "result_verified": true, "effect": null, "ci": "dU +0.0003 (−0.019..+0.020)", "mde": null, "verdict": "closed", "verdict_detail": "FAIL (neutral swap)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1936" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-269", "code": "CITESUP", "name": "CITESUP", "family": "models", "question": "Embedding-cosine rescue of dropped citations", "hypothesis": null, "preregistered": { "status": "yes (127f17db)", "where_public": null }, "set": "2 screens", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "recovers ~1/3 of dropped citations; Ministral still NO-GO", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1937" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-270", "code": "GEMMA-VS-4B", "name": "GEMMA-VS-4B", "family": "models", "question": "How much of the 4B gain does Gemma capture?", "hypothesis": null, "preregistered": { "status": "yes (829e3ba1)", "where_public": null }, "set": "MAIN 97 K0", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "U 1.7B 0.406, Gemma 0.434, 4B 0.481; Gemma captures 38 %", "result_verified": true, "effect": null, "ci": "4B−1.7B +0.076 (+0.050..+0.101); 4B−Gemma +0.048 (+0.021..+0.075)", "mde": null, "verdict": "measurement", "verdict_detail": "measurement", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1938" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-271", "code": "GEMMA-PACKS", "name": "GEMMA-PACKS", "family": "packs-delivery", "question": "Does Gemma read packs better than the 1.7B?", "hypothesis": null, "preregistered": { "status": "yes (e4c9a4eb)", "where_public": null }, "set": "80 pack-matched self-tests (pack-written)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "Gemma K3−K0 +0.102 vs 1.7B +0.069", "result_verified": true, "effect": null, "ci": "DiD +0.034 (+0.001..+0.066)", "mde": null, "verdict": "not-detected", "verdict_detail": "measurement (by a hair)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1939" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-272", "code": "INPACK-SPLIT", "name": "INPACK-SPLIT", "family": "packs-delivery", "question": "Are in-pack items easier; does the matching pack help on real questions?", "hypothesis": null, "preregistered": { "status": "post hoc (0 LLM)", "where_public": null }, "set": "MAIN 97; 17 matched real items; 60 self-tests", "judge": "existing Claude Haiku", "n": null, "seeds": null, "result": "real matched K3−K0 −0.023 (−0.081..+0.027); self-tests +0.056 (+0.018..+0.096)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "pack gain only on self-tests", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1940" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-273", "code": "INPACK-STRATUM", "name": "INPACK-STRATUM", "family": "submitted-app", "question": "Which v0.4 items are in-pack under A∪?", "hypothesis": null, "preregistered": { "status": "rule INPACK-R1 (not frozen then)", "where_public": null }, "set": "v0.4 tags only (held-out, ids/tags read by script)", "judge": "rule", "n": null, "seeds": null, "result": "57/75 in-pack under A∪ (55/16 in the 71 headline); 0/75 under the native kit", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "stratum defined", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1941" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24" }, { "id": "X-274", "code": "FACT-REACH", "name": "FACT-REACH", "family": null, "question": "Where do the 20 no-source CEs' key facts live?", "hypothesis": null, "preregistered": { "status": "no (0 LLM)", "where_public": null }, "set": "20 dev items", "judge": "read-only FTS", "n": null, "seeds": null, "result": "top100k 7+2 partial; full enwiki 10+7; published/wave-4 16+4; partly true by construction", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "routing/retrieval/reading bind", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1942" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-275", "code": "WIKIVOYAGE-REBUILD", "name": "WIKIVOYAGE-REBUILD", "family": "packs-delivery", "question": "Template whitelist for wikivoyage", "hypothesis": null, "preregistered": { "status": "yes (1032ab28)", "where_public": null }, "set": "pack build + 280 replays", "judge": "holes", "n": null, "seeds": null, "result": "fact holes 1.36→0.79 %; literal verdict INVALID by a precondition (27 chunks)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "candidate, not shipped", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1943" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-276", "code": "Q17B73-PHONE", "name": "Q17B73-PHONE", "family": "models", "question": "1.7B on the same 73 and phone as Gemma", "hypothesis": null, "preregistered": { "status": "no (measurement)", "where_public": null }, "set": "73 phone items", "judge": "timing", "n": null, "seeds": null, "result": "p50 4.2 s, p90 12.2 s, 12.4 tok/s, PSS 3.39 GB", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "measurement", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1944" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-277", "code": "HON-1 WARNCAL", "name": "HON-1 WARNCAL", "family": null, "question": "Is the warning line precise?", "hypothesis": null, "preregistered": { "status": "yes (d641551b)", "where_public": null }, "set": "746 non-decline answers (dev)", "judge": "existing Claude Haiku", "n": null, "seeds": null, "result": "precision 0.39 (0.32-0.46), recall 0.67 (0.59-0.75); Opus 0.40 / 0.65", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "not-detected", "verdict_detail": "kill not met by a hair", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1945" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-278", "code": "HON-2 SHOWANYWAY", "name": "HON-2 SHOWANYWAY", "family": null, "question": "Is \"show anyway\" worth it?", "hypothesis": null, "preregistered": { "status": "yes (same)", "where_public": null }, "set": "492 revealable hidden answers", "judge": "Claude Haiku; Claude Opus (WARN-OPUS, REVEAL-CENSUS)", "n": null, "seeds": null, "result": "Haiku correct:CE 0.19 (0.12-0.31); Opus 0.54 (0.32-0.83) on pool; MAIN census 11:16 = 0.69", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "liability as specified; second confirmation, non-health only", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1946" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-279", "code": "DIVERGENCE-MAP", "name": "DIVERGENCE-MAP", "family": "ruler", "question": "How many items are deterministic across systems?", "hypothesis": null, "preregistered": { "status": "no (0 calls)", "where_public": null }, "set": "23 judged conditions", "judge": "counts", "n": null, "seeds": null, "result": "30/72 phone v0.1 ids identical text across 1.7B/Gemma/4B; per-item seed r 0.66", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "divergence-first rule", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1947" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-280", "code": "AUNION-PIXEL", "name": "AUNION-PIXEL", "family": "packs-delivery", "question": "Six A∪ packs on Pixel 8a with 1.7B / Gemma / 4B", "hypothesis": null, "preregistered": { "status": "yes (a852f6bd + addenda)", "where_public": null }, "set": "73 phone items", "judge": "timing / PSS", "n": null, "seeds": null, "result": "1.7B answer p50 7.6 vs 4.2 s, PSS 3.41 GB, 0 kills; Gemma PSS 4.38 GB; 4B PSS 4.76 GB; cost = +2.4 s sequential pack search", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "PASS / MARGINAL (no memory cost)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1948" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24" }, { "id": "X-281", "code": "FACTPRESENT", "name": "FACTPRESENT", "family": "packs-delivery", "question": "Does the self-test pack gain sit on items whose fact the pack adds, and does it hold on real items?", "hypothesis": null, "preregistered": { "status": "yes (rule frozen first)", "where_public": null }, "set": "60 topical self-tests; MAIN 92 keyed", "judge": "existing Claude Haiku", "n": null, "seeds": null, "result": "self-test cell a +0.139 (+0.075..+0.201) n 27; real cell a −0.010 (−0.077..+0.050) n 10", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "gain does not reproduce on real items", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1949" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-282", "code": "SPEED-PRECHECK", "name": "SPEED-PRECHECK", "family": "ruler", "question": "Are the tiers usable by the ruler's speed rule?", "hypothesis": null, "preregistered": { "status": "yes (4bd52b3d)", "where_public": null }, "set": "Pixel 8a traces", "judge": "T30 / complete", "n": null, "seeds": null, "result": "all four USABLE; 1.7B+A∪ p50 T30 4.4 s; 4B p50 T30 20.2 s, complete 26.0 s", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "4B also passes", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1950" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24" }, { "id": "X-283", "code": "FLOOR-ANCHORS", "name": "FLOOR-ANCHORS", "family": "ruler", "question": "Does judge.v1 give a true zero to declines; is the 1.7B above a closed-book 1B-class model?", "hypothesis": null, "preregistered": { "status": "yes (ed88f377)", "where_public": null }, "set": "30 + 57 v0.1 (dev)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "bare refusal U 0.058 (0.037-0.081), hedge 0.053, fluent no-content 0.016; 1.5B closed-book 0.259 vs 1.7B 0.294", "result_verified": true, "effect": null, "ci": "1.7B−1.5B +0.035 (−0.019..+0.089)", "mde": null, "verdict": "closed", "verdict_detail": "decline arm INVALID; 1B not separable", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1951" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-284", "code": "TILE-PATH", "name": "TILE-PATH", "family": "submitted-app", "question": "Can the current-city path be scored on the apps?", "hypothesis": null, "preregistered": { "status": "no (analysis)", "where_public": null }, "set": null, "judge": "code read", "n": null, "seeds": null, "result": "8,374 tiles hosted; the native app has no places or location code", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "stratum scored as given", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1952" ], "chapter": "submission", "source": "verify pass 2026-10-24)" }, { "id": "X-285", "code": "REFVAR", "name": "REFVAR", "family": null, "question": "Does reference draw or length change U?", "hypothesis": null, "preregistered": { "status": "yes (bc368791)", "where_public": null }, "set": "20 v0.1 (dev), 2 uncapped Opus+web draws", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "draw A−B moves U 0.015 / 0.008; uncapped vs v1 +0.013 / +0.001; no flip", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "one draw per family suffices", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1953" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-286", "code": "RAG-FUNNEL", "name": "RAG-FUNNEL", "family": "packs-delivery", "question": "Where are facts lost (S1 not in pack → S5)?", "hypothesis": null, "preregistered": { "status": "no (0 LLM replay)", "where_public": null }, "set": "MAIN 92 keyed", "judge": "stage counts", "n": null, "seeds": null, "result": "K3: S1 45 (39 cooking/money), S2 7, S3 dropped at keep 17, S4 kept-wrong 15, S5 8", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "diagnosis", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1954" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-287", "code": "JUDGE-V1_1 / V1_1B", "name": "JUDGE-V1_1 / V1_1B", "family": "ruler", "question": "Can one rule line make judge.v1 give declines a true zero?", "hypothesis": null, "preregistered": { "status": "yes (bf560ea9", "where_public": null }, "set": "decline variants; 40 controls", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "declines U 0.000, but CE flips 7/40 and 5/40; retest SD 0.0265/0.0295 vs 0.0206", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-ADOPT both", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1955" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-288", "code": "COVERAGE-MATRIX", "name": "COVERAGE-MATRIX", "family": "ruler", "question": "Do packs help all three tiers on self-tests; does A∪ harm MAIN?", "hypothesis": null, "preregistered": { "status": "yes (dfc84677)", "where_public": null }, "set": "80 self-tests (pack-written); MAIN 97", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "KA vs K0 +0.052..+0.055 all tiers; 4B−1.7B +0.08 at every pack set; MAIN 1.7B KA−K0 −0.007 (−0.026..+0.012), CE-unwarned 7→12", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "upper bound; no harm shown on U", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1956" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-289", "code": "DRIFT17", "name": "DRIFT17", "family": "ruler", "question": "Does the 1.7B degrade over a session?", "hypothesis": null, "preregistered": { "status": "yes (5ea4b345)", "where_public": null }, "set": "longsession-v1 80 (dev)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "CE-unwarned t1 0/10 vs t4-5 3/20, CI covers 0; one unsafe health error in both arms", "result_verified": true, "effect": null, "ci": "rise CI 0.00..0.30", "mde": null, "verdict": "not-detected", "verdict_detail": "KILLED (no drift detected)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1957" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-290", "code": "ORACLE-REAL", "name": "ORACLE-REAL", "family": "packs-delivery", "question": "Does forcing the fact chunk to passage 1 help on real items?", "hypothesis": null, "preregistered": { "status": "yes (790fe1d0)", "where_public": null }, "set": "30 keyed v0.1 (dev)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "U 0.427 vs 0.412", "result_verified": true, "effect": null, "ci": "+0.015 (−0.029..+0.059)", "mde": null, "verdict": "closed", "verdict_detail": "KILL (reading binds)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1958" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-291", "code": "RV2-SEP-LADDER", "name": "RV2-SEP-LADDER", "family": "ruler", "question": "Does the ruler separate tiers?", "hypothesis": null, "preregistered": { "status": "yes (53f75edc)", "where_public": null }, "set": "v0.1 51 (dev)", "judge": "existing Claude Haiku", "n": null, "seeds": null, "result": "1.5B CB 0.209, 1.7B 0.249, Gemma 0.260, 4B 0.376; coverage DiD fails", "result_verified": true, "effect": null, "ci": "4B−Gemma +0.117 (+0.074..+0.162); 1.7B−1.5B +0.040 (−0.019..+0.096)", "mde": null, "verdict": "measurement", "verdict_detail": "4B separable; 1.7B ≈ closed-book 1.5B (Haiku)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1959" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-292", "code": "RV4-CITY-BOUND", "name": "RV4-CITY-BOUND", "family": "ruler", "question": "How much can the 7 current-city items swing v0.4?", "hypothesis": null, "preregistered": { "status": "yes (53f75edc)", "where_public": null }, "set": "v0.4 tags (held-out, script)", "judge": "arithmetic", "n": null, "seeds": null, "result": "7 items = 10.7 % of headline weight; swing ≥ 0.107 U", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "KILL branch: separate stratum scored as given", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1960" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-293", "code": "HON-7 WARN-TRANSFER", "name": "HON-7 WARN-TRANSFER", "family": "models", "question": "Does the warning line work for Gemma and the 4B?", "hypothesis": null, "preregistered": { "status": "yes (f53c867c)", "where_public": null }, "set": "MAIN 97", "judge": "existing Claude Haiku", "n": null, "seeds": null, "result": "Gemma, 1.7B HOLD; 4B precision 0.275 (0.15-0.42) INDETERMINATE", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "refuse-by-default validated for 1.7B/Gemma only", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1961" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-294", "code": "HON-6 / CV5-PER100-OPUS", "name": "HON-6 / CV5-PER100-OPUS", "family": "models", "question": "Per 100 questions, refuse-by-default: correct shown vs CE shown by tier", "hypothesis": null, "preregistered": { "status": "yes (f53c867c", "where_public": null }, "set": "MAIN 97", "judge": "Claude Haiku; Claude Opus", "n": null, "seeds": null, "result": "Haiku tie 1.7B=4B 16.5 correct; Opus 4B 22.7, Gemma 23.7, 1.7B 18.6; CE shown 1.0 / 0.0 / 2.1", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "MIXED (Haiku); tie KILLED (Opus)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1962" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-295", "code": "GEMMA-SAFETY", "name": "GEMMA-SAFETY", "family": "safety", "question": "Is Gemma safe behind the warning stack?", "hypothesis": null, "preregistered": { "status": "yes (1c0389ba)", "where_public": null }, "set": "65 safety probes", "judge": "blinded agent reading", "n": null, "seeds": null, "result": "Gemma unsafe 1 vs 1.7B 2", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "PASS", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1963" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-296", "code": "PK-5 MERGED-PACK", "name": "PK-5 MERGED-PACK", "family": "ruler", "question": "One merged A∪ pack instead of six", "hypothesis": null, "preregistered": { "status": "yes (59a909f0 + addendum)", "where_public": null }, "set": "MAIN 97 replay; 16 changed items", "judge": "identity; Claude Haiku", "n": null, "seeds": null, "result": "kept identity 83.5 % < 95 %; judged U 0.365 vs 0.399", "result_verified": true, "effect": null, "ci": "−0.034 (< −0.02)", "mde": null, "verdict": "closed", "verdict_detail": "FAIL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1964" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-297", "code": "PK-3 OVERLAP", "name": "PK-3 OVERLAP", "family": "packs-delivery", "question": "Is the self-test pack gain a lexical-overlap artefact?", "hypothesis": null, "preregistered": { "status": "yes (fe44d05e)", "where_public": null }, "set": "cand80 vs 37 real golds", "judge": "reweighting", "n": null, "seeds": null, "result": "reweighted to real overlap +0.064 (+0.035..+0.095)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "KILL (not an artefact; question type/reading)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1965" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-298", "code": "PK-4 PACK-PRUNE", "name": "PK-4 PACK-PRUNE", "family": "packs-delivery", "question": "Can any A∪ pack be pruned?", "hypothesis": null, "preregistered": { "status": "yes (fe44d05e)", "where_public": null }, "set": "280-item replay", "judge": "kept counts", "n": null, "seeds": null, "result": "every pack reaches kept context (medlineplus 26 items)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "KILL (keep six)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1966" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-299", "code": "RV1-DECLINE-CU", "name": "RV1-DECLINE-CU", "family": "ruler", "question": "U from completeness+usefulness only", "hypothesis": null, "preregistered": { "status": "yes (b0aa1ac3)", "where_public": null }, "set": "v0.1 51", "judge": "recompute", "n": null, "seeds": null, "result": "declines 0, retest ok, but 4B−1.5B gap moves −0.039 (−0.062..−0.015)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-ADOPT", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1967" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-300", "code": "HON-8 REVEAL-RULE", "name": "HON-8 REVEAL-RULE", "family": null, "question": "Is any subset of hidden answers safe to reveal?", "hypothesis": null, "preregistered": { "status": "yes (b0aa1ac3)", "where_public": null }, "set": "492 revealable rows", "judge": "counts", "n": null, "seeds": null, "result": "none reaches correct:CE ≥ 1.0; best 0.37", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "KILL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1968" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-301", "code": "RD-1 CONVERSION-MAP", "name": "RD-1 CONVERSION-MAP", "family": "packs-delivery", "question": "Does chunk form explain why the 1.7B ignores a present fact?", "hypothesis": null, "preregistered": { "status": "yes (84c2c5a2)", "where_public": null }, "set": "fact-present rows", "judge": "existing Claude Haiku", "n": null, "seeds": null, "result": "conversion: 1.7B 0.33, Gemma 0.72, 4B 0.57; best ceiling 0.002 U", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "KILL (reading capacity)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1969" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-302", "code": "CHUNKCLEAN", "name": "CHUNKCLEAN", "family": "packs-delivery", "question": "Hand-cleaned fact sentence vs raw chunk", "hypothesis": null, "preregistered": { "status": "yes (10f3c2df)", "where_public": null }, "set": "30 real v0.1", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "C−R +0.031", "result_verified": true, "effect": null, "ci": "+0.031 (−0.009..+0.073)", "mde": null, "verdict": "not-detected", "verdict_detail": "INCONCLUSIVE", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1970" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-303", "code": "CV5-LADDER-OPUS", "name": "CV5-LADDER-OPUS", "family": "ruler", "question": "Ladder under Opus", "hypothesis": null, "preregistered": { "status": "yes (3870ac70)", "where_public": null }, "set": "v0.1 51", "judge": "Claude Opus, 324 calls", "n": null, "seeds": null, "result": "4B−1.7B +0.109 (+0.063..+0.160); 4B−Gemma +0.098; 1.7B−1.5B CB +0.090 (+0.009..+0.170) vs Haiku +0.038", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "4B gap holds; \"1.7B level with 1B-class\" sentence killed", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1971" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24" }, { "id": "X-304", "code": "RD-4 DILUTION-K", "name": "RD-4 DILUTION-K", "family": null, "question": "Gold-only context vs gold + rest", "hypothesis": null, "preregistered": { "status": "yes (af00dcaa)", "where_public": null }, "set": "30 ORACLE-REAL items", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "k1 − R −0.009", "result_verified": true, "effect": null, "ci": "−0.009 (−0.049..+0.033)", "mde": null, "verdict": "closed", "verdict_detail": "KILL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1972" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-305", "code": "CV5-RECOMPUTE / EXPOSURE-AUDIT", "name": "CV5-RECOMPUTE / EXPOSURE-AUDIT", "family": "ruler", "question": "Do page numbers re-derive; are v0.4 and held-out ledgers clean?", "hypothesis": null, "preregistered": { "status": "yes (78f43da6)", "where_public": null }, "set": "295 page numbers; 2,203 run files", "judge": "script", "n": null, "seeds": null, "result": "263/270 re-derive (7 mismatches, 2.6 %); 25/295 unsourced; v0.4 0 run files; 5 missing ledger rows", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "PASS with one label fix", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1973" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-306", "code": "WARN-TUNE", "name": "WARN-TUNE", "family": null, "question": "Can a trace-feature rule lift warning precision to 0.55?", "hypothesis": null, "preregistered": { "status": "yes (1799c624)", "where_public": null }, "set": "fit 180, test MAIN 97", "judge": "existing labels", "n": null, "seeds": null, "result": "goes the wrong way on MAIN (0.407→0.383)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "KILL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1974" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-307", "code": "FEWSHOT-READ", "name": "FEWSHOT-READ", "family": "prompting-style", "question": "Two worked examples in the 1.7B prompt", "hypothesis": null, "preregistered": { "status": "yes (081c410c)", "where_public": null }, "set": "stage1-pool 87", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "U 0.396→0.395", "result_verified": true, "effect": null, "ci": "−0.001 (−0.015..+0.014)", "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1975" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-308", "code": "SAMPLE2-GATE", "name": "SAMPLE2-GATE", "family": null, "question": "Best-of-2 seeds picked by the gate", "hypothesis": null, "preregistered": { "status": "yes (081c410c)", "where_public": null }, "set": "stage1-pool 87", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "decode x2", "result_verified": true, "effect": null, "ci": "+0.003 (−0.007..+0.014)", "mde": null, "verdict": "closed", "verdict_detail": "NO-GO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1976" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-309", "code": "HAIKU-CAL", "name": "HAIKU-CAL", "family": null, "question": "Map Haiku-low to Opus on double-labelled units", "hypothesis": null, "preregistered": { "status": "yes (spec frozen first)", "where_public": null }, "set": "1,911 units", "judge": "κ", "n": null, "seeds": null, "result": "κ CE 0.75, correct 0.49 (warned 0.22); U shift Opus−Haiku +0.022", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "calibration map", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1977" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-310", "code": "DECLINE2STAGE", "name": "DECLINE2STAGE", "family": "ruler", "question": "A Haiku stage-1 non-answer classifier before judge.v1", "hypothesis": null, "preregistered": { "status": "yes (43d52938)", "where_public": null }, "set": "FLOOR declines 90; MAIN 49; WARN 27", "judge": "Claude Haiku, 258 calls", "n": null, "seeds": null, "result": "recall 90/90; bar (b) 2 of 49 zeroed (both the engine refusal string) vs cap 1", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "FAIL by letter → RULING R1 adopted", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1978" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24" }, { "id": "X-311", "code": "X2B-RETEST / X2B-OFF-MAIN", "name": "X2B-RETEST / X2B-OFF-MAIN", "family": "ruler", "question": "Is the X2b head pass worth its tokens on real items?", "hypothesis": null, "preregistered": { "status": "yes (3a3aaf0b", "where_public": null }, "set": "stage1-pool 87; MAIN 77 live, 2 seeds", "judge": "Claude Haiku; Claude Opus on flips", "n": null, "seeds": null, "result": "real +0.033 (−0.100..+0.173) n 25; OFF pooled dU_d +0.026 (−0.020..+0.073), CE-unwarned +2; prompt −42 %", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "not-detected", "verdict_detail": "underpowered; OFF safe only at edge (kit-only)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1979" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-312", "code": "HYBRID-READER", "name": "HYBRID-READER", "family": "packs-delivery", "question": "Route hard items to the 4B reader by retrieval features", "hypothesis": null, "preregistered": { "status": "yes (87060c26)", "where_public": null }, "set": "MAIN 97", "judge": "existing Claude Haiku/Claude Opus", "n": null, "seeds": null, "result": "routes 5/97, +0.004; random routing similar", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "KILL (chance level)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1980" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-313", "code": "ROUTING-R1", "name": "ROUTING-R1", "family": "safety", "question": "Fix routing losses", "hypothesis": null, "preregistered": { "status": "yes (0fff7357)", "where_public": null }, "set": "MAIN 97 + 64 probes", "judge": "replay + Claude Haiku", "n": null, "seeds": null, "result": "7/97 losses; fixes drop safety probes off the health path (up to 32/64)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "none adopted", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1981" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-314", "code": "KEEP-K1 / KEEP-E", "name": "KEEP-K1 / KEEP-E", "family": "packs-delivery", "question": "Keep-step variants to bring dropped fact chunks in", "hypothesis": null, "preregistered": { "status": "yes (keep-e 4421544c)", "where_public": null }, "set": "MAIN 97 + 180 replays; 14 changed", "judge": "0 LLM; Claude Haiku", "n": null, "seeds": null, "result": "best 7/17 at +50 % tokens; KEEP-E net U −0.010", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "CLOSED / KILL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1982" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-315", "code": "E1-PARTIME / E1-PAR", "name": "E1-PARTIME / E1-PAR", "family": "packs-delivery", "question": "Parallel pack search on the bench engine", "hypothesis": null, "preregistered": { "status": "yes (ffd0d931)", "where_public": null }, "set": "phone traces; 10 phone items per arm", "judge": "timing", "n": null, "seeds": null, "result": "projection 7.57→5.5-6.0 s; device: search 3.89→1.93 s, answer 10.32→7.39 s; cache-64 arm KILL by rule", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "Promise.all PAYS (bench engine only)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1983" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-316", "code": "NOTE2-TRANSFER", "name": "NOTE2-TRANSFER", "family": "models", "question": "Do desktop hide rates transfer to the phone?", "hypothesis": null, "preregistered": { "status": "yes (b6b1f787)", "where_public": null }, "set": "57 shared v0.1", "judge": "counts", "n": null, "seeds": null, "result": "1.7B A∪ phone vs desktop +26 pp", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "KILLED", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1984" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-317", "code": "BLOCKCACHE", "name": "BLOCKCACHE", "family": "packs-delivery", "question": "Decompressed-block LRU for pack search", "hypothesis": null, "preregistered": { "status": "yes (4dab635a)", "where_public": null }, "set": "session replay", "judge": "counts", "n": null, "seeds": null, "result": "256 slots −78 % decompressions but +65 MB > 32 MB", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "CLOSED as frozen", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1985" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-318", "code": "DEFAULT-SET", "name": "DEFAULT-SET", "family": "packs-delivery", "question": "Which A∪ packs to install by default", "hypothesis": null, "preregistered": { "status": "yes (2c8e3010)", "where_public": null }, "set": "112 ids + 64 probes", "judge": "replay", "n": null, "seeds": null, "result": "frozen clause keeps visa and medlineplus; options table for author", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "KEEP", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1986" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24)" }, { "id": "X-319", "code": "C1-REASON", "name": "C1-REASON", "family": "packs-delivery", "question": "Why are 16 owned-pack facts missed?", "hypothesis": null, "preregistered": { "status": "no (0 LLM)", "where_public": null }, "set": "16 facts", "judge": "replay", "n": null, "seeds": null, "result": "7 retrieval, 4 routing, 2 build, 2 install-only, 1 reading; install reaches 2/9", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "C1 closes", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1987" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-320", "code": "FA-CARDS", "name": "FA-CARDS", "family": null, "question": "Are there coverable first-aid card misses?", "hypothesis": null, "preregistered": { "status": "yes (42b910a3)", "where_public": null }, "set": "245 rows", "judge": "replay", "n": null, "seeds": null, "result": "0 beyond epinephrine; trigger edit routes 0/2 epi turns", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "KILLED", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1988" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-321", "code": "CONVERT-FIX", "name": "CONVERT-FIX", "family": "packs-delivery", "question": "Converter drops unit-conversion values", "hypothesis": null, "preregistered": { "status": "n/a (bug + 11 tests)", "where_public": null }, "set": "pack rebuild", "judge": "counts", "n": null, "seeds": null, "result": "artefact chunks 3,176→524; simplewiki r2 published 15:01", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "shipped (pack r2)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1989" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-322", "code": "ZSTD-BENCH / PARTIAL-DECOMP / REPACK16", "name": "ZSTD-BENCH / PARTIAL-DECOMP / REPACK16", "family": "packs-delivery", "question": "Faster pack block decode", "hypothesis": null, "preregistered": { "status": "yes (REPACK16 daedb7dd)", "where_public": null }, "set": "200 blocks; replays", "judge": "timing", "n": null, "seeds": null, "result": "WASM zstd 3-4x faster (no WASM on Hermes, unresolved); 16 KB repack −18 % bytes at 64 slots (rule 50 %)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "REPACK16 KILL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1990" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-324", "code": "A1-SUCCINCT17", "name": "A1-SUCCINCT17", "family": "submitted-app", "question": "Time saved by Succinct for small models", "hypothesis": null, "preregistered": { "status": "yes (d037c010)", "where_public": null }, "set": "56 v0.1 native exports", "judge": "tokens × tok/s", "n": null, "seeds": null, "result": "Detailed +348 tokens p50 → ~24 s per answer at 14.3 tok/s (projection)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "SUPPORTED pending device", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1992" ], "chapter": "submission", "source": "verify pass 2026-10-24)" }, { "id": "X-325", "code": "A4-CAP512", "name": "A4-CAP512", "family": "prompting-style", "question": "Is a 512-token Detailed cap a lever?", "hypothesis": null, "preregistered": { "status": "yes (a1d6bf6e)", "where_public": null }, "set": "404 Detailed answers", "judge": "tokens", "n": null, "seeds": null, "result": "saves 0 s p50; truncates 10.9 %", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NOT A LEVER", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1993" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-326", "code": "LENGTH-SLOPE", "name": "LENGTH-SLOPE", "family": "ruler", "question": "Does judge.v1 reward length?", "hypothesis": null, "preregistered": { "status": "no (0 LLM, 13,310 rows)", "where_public": null }, "set": "existing rows", "judge": "within-item slope", "n": null, "seeds": null, "result": "no length reward except the 1.7B under Opus (+0.31 per doubling)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "Succinct not penalised at ≥4B", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1994" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-327", "code": "G1-SWA5", "name": "G1-SWA5", "family": "models", "question": "Gemma 4 sliding-window KV bug in 5-turn sessions on the phone runtime", "hypothesis": null, "preregistered": { "status": "yes (b89f1f1a)", "where_public": null }, "set": "10 turns per arm, Pixel 8a", "judge": "markers", "n": null, "seeds": null, "result": "0 decode/KV markers, 0 garbled", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "KILLED (no bug seen)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1995" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-329", "code": "SHIP15B-PACKS", "name": "SHIP15B-PACKS", "family": "packs-delivery", "question": "Do packs help the 1.5B the app ships?", "hypothesis": null, "preregistered": { "status": "yes (4f9e7d5a)", "where_public": null }, "set": "v0.1 51", "judge": "Claude Opus + D2S; Claude Haiku", "n": null, "seeds": null, "result": "0.270 vs 0.255; CE shown 25.5 vs 31.4 per 100", "result_verified": true, "effect": null, "ci": "+0.015 (−0.036..+0.068)", "mde": null, "verdict": "closed", "verdict_detail": "NO-GAIN", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1997" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-330", "code": "STYLE15B", "name": "STYLE15B", "family": "prompting-style", "question": "Succinct vs Detailed on 1.5B and 1.7B", "hypothesis": null, "preregistered": { "status": "yes (474ac0a0)", "where_public": null }, "set": "v0.1 51", "judge": "Claude Haiku; Claude Opus diagnostic 11", "n": null, "seeds": null, "result": "1.5B 0.220 vs 0.133; 1.7B 0.263 vs 0.224; CE shown halves", "result_verified": true, "effect": null, "ci": "1.5B +0.087 (+0.038..+0.143); 1.7B +0.038 (−0.005..+0.084)", "mde": null, "verdict": "parity", "verdict_detail": "Succinct NONINFERIOR", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-1998" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-332", "code": "CB17", "name": "CB17", "family": "packs-delivery", "question": "Do packs move the 1.7B over closed-book?", "hypothesis": null, "preregistered": { "status": "criterion in arm", "where_public": null }, "set": "MAIN 97; v0.1 51", "judge": "Claude Opus 58; Claude Haiku 291", "n": null, "seeds": null, "result": "CB 0.337 vs K0 0.351 (Opus 51); 63/97 texts identical", "result_verified": true, "effect": null, "ci": "K0−CB +0.015 (−0.035..+0.056)", "mde": null, "verdict": "closed", "verdict_detail": "packs not separable on aggregate", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-2000" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-333", "code": "LA1 / LA5", "name": "LA1 / LA5", "family": null, "question": "Can section 0 state app and engine numbers without misattribution?", "hypothesis": null, "preregistered": { "status": "yes (49d1c5ac", "where_public": null }, "set": "text variants", "judge": "1 Claude Opus reviewer + 3 cold readers", "n": null, "seeds": null, "result": "LA1 0 misattribution; LA5 withhold = material omission", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "LA1 PASS; LA5 KILLED", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-2001" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-334", "code": "C1-GAP-LEDGER", "name": "C1-GAP-LEDGER", "family": "packs-delivery", "question": "Sourced cause ledger for the app gap", "hypothesis": null, "preregistered": { "status": "ledger sha 3e233f6e", "where_public": null }, "set": "code + device files", "judge": "checker", "n": null, "seeds": null, "result": "5 rows verified (no 1.7B in catalog; Detailed hook; 0 packs; no warning text; KB cutoff 0.7); layers not established", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "ledger", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-2002" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-338", "code": "NOREPACK97", "name": "NOREPACK97", "family": "packs-delivery", "question": "Cross-host determinism flag", "hypothesis": null, "preregistered": { "status": "yes (27d05cde)", "where_public": null }, "set": "MAIN 97", "judge": "identity; Claude Haiku", "n": null, "seeds": null, "result": "97/97 identical across hosts", "result_verified": true, "effect": null, "ci": "dU −0.012 (−0.030..+0.005)", "mde": null, "verdict": "adopted", "verdict_detail": "adopted for pairing only", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-2006" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-339", "code": "RC1-STAGE1", "name": "RC1-STAGE1", "family": "submitted-app", "question": "Does stage 1 wrongly zero long app answers?", "hypothesis": null, "preregistered": { "status": "yes (824d94f5)", "where_public": null }, "set": "51 RC1 answers", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "0 of 51 marked non-answer", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "PASS (low power)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-2007" ], "chapter": "submission", "source": "verify pass 2026-10-24)" }, { "id": "X-340", "code": "FACT-AVAIL", "name": "FACT-AVAIL", "family": "ruler", "question": "Which v0.4 items have their answer in the A∪ packs?", "hypothesis": null, "preregistered": { "status": "yes (1ac3fa96 + addendum)", "where_public": null }, "set": "v0.4 71 headline (held-out; labels quarantined)", "judge": "Claude Haiku 329 + Claude Opus 108 labels", "n": null, "seeds": null, "result": "AVAILABLE 19 / PARTIAL 25 / NONE 27 (strict 15/23/33); validation 34/40", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "label set (quarantined)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-2008" ], "chapter": "new-ruler", "source": "verify pass 2026-10-24" }, { "id": "X-341", "code": "DECODE-GROUND", "name": "DECODE-GROUND", "family": null, "question": "Grounding at the decoder (grammar select, context-aware decoding, closed-book fallback)", "hypothesis": null, "preregistered": { "status": "yes (92d3340a", "where_public": null }, "set": "GATE v2 K1 screen; MAIN 97", "judge": "Claude Haiku 276", "n": null, "seeds": null, "result": "all arms net 4-6 (bar 9)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "CLOSED (null)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-2009" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-342", "code": "QFORM", "name": "QFORM", "family": "packs-delivery", "question": "Does question phrasing decide whether the pack chunk is found, and how well does the 1.7B answer when it is?", "hypothesis": null, "preregistered": { "status": "yes (8371f509)", "where_public": null }, "set": "48 pack facts × 9 forms = 432 (dev, agent-written)", "judge": "replay; Claude Haiku 384", "n": null, "seeds": null, "result": "hit@kept base 50 %, entity-free 10 %, follow-up 8 %, situational 25 %; reference kept U 0.88 (CE 2/43) vs not 0.19 (CE 75/147)", "result_verified": true, "effect": null, "ci": "F1 −40 (−54..−23 pts), F8 −42, F3 −25", "mde": null, "verdict": "measurement", "verdict_detail": "form is a retrieval effect", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-2010" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-343", "code": "QFORM-W74", "name": "QFORM-W74", "family": "packs-delivery", "question": "Previous turn + fragment as the search query", "hypothesis": null, "preregistered": { "status": "no (0 LLM)", "where_public": null }, "set": "48 follow-up fragments", "judge": "replay", "n": null, "seeds": null, "result": "kept 4/48 → 24/48; the app's own follow-up rule already gives 20/48", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "bench-engine artefact", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-21", "fact_ids": [ "F-2011" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-344", "code": "RETRIEVE-SWEEP", "name": "RETRIEVE-SWEEP", "family": "packs-delivery", "question": "Can engine knobs guarantee the chunk is found?", "hypothesis": null, "preregistered": { "status": "yes (94cde143)", "where_public": null }, "set": "QFORM 432; funnel 25/22", "judge": "replay", "n": null, "seeds": null, "result": "best sibling arm +7.9 pts (junk 2.4x); real recovery ≤ 2/25", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "no arm passes", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2012" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-345", "code": "ARM-G chunk vectors", "name": "ARM-G chunk vectors", "family": "packs-delivery", "question": "Chunk-level vectors for packs", "hypothesis": null, "preregistered": { "status": "criterion in arm (G2 dcac6f4b)", "where_public": null }, "set": "216 QFORM items; MAIN 97", "judge": "replay; Claude Haiku 336-400", "n": null, "seeds": null, "result": "kept 21→32-36 %; answers +0.109 (+0.026..+0.190), final +0.125 (+0.049..+0.204) on pack items; whole kit real recovery 3/25", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "pack items yes; G2 FAIL on real items", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2013" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-346", "code": "T1-TERMS", "name": "T1-TERMS", "family": "packs-delivery", "question": "Model-free search-term extraction for situational questions", "hypothesis": null, "preregistered": { "status": "yes (b7d73bf0)", "where_public": null }, "set": "QFORM 432", "judge": "replay", "n": null, "seeds": null, "result": "situational 25 %→23 %", "result_verified": true, "effect": null, "ci": "−2 (−15..+10)", "mde": null, "verdict": "closed", "verdict_detail": "FAIL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2014" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-347", "code": "COVERAGE-MAP", "name": "COVERAGE-MAP", "family": "packs-delivery", "question": "Are missing dev facts absent from packs on disk?", "hypothesis": null, "preregistered": { "status": "no (0 LLM)", "where_public": null }, "set": "MAIN 92 keyed", "judge": "replay", "n": null, "seeds": null, "result": "installing 18 packs moves S1 52→4 but kept only 23→28", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "after install the gap is retrieval", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2015" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-348", "code": "SEARCH-AUTOPSY / H1-H3", "name": "SEARCH-AUTOPSY / H1-H3", "family": "packs-delivery", "question": "Autopsy of 25 lost real items and 3 targeted fixes", "hypothesis": null, "preregistered": { "status": "yes (94cde143 rule)", "where_public": null }, "set": "MAIN 97", "judge": "replay; Claude Haiku 150", "n": null, "seeds": null, "result": "only ~9 are retrieval losses; H1+H2 answer screen −0.023 (−0.052..+0.006)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "FAIL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2016" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-349", "code": "B3-FULLWIKI", "name": "B3-FULLWIKI", "family": null, "question": "Do full-Wikipedia shards deliver facts?", "hypothesis": null, "preregistered": { "status": "criterion in arm", "where_public": null }, "set": "MAIN 97 × 15 shards", "judge": "replay", "n": null, "seeds": null, "result": "each shard covers 1-5 more facts, kept −3..+3", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "COVERAGE-ONLY (scored with old merge)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2017" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-350", "code": "H4-MERGE", "name": "H4-MERGE (gidf)", "family": "packs-delivery", "question": "Global-IDF re-scoring across packs", "hypothesis": null, "preregistered": { "status": "criterion in arm", "where_public": null }, "set": "QFORM six separate packs; MAIN 92", "judge": "replay", "n": null, "seeds": null, "result": "+6.2 pts kept (27 gained, 0 lost, junk 1.10x); rrf worse", "result_verified": true, "effect": null, "ci": "+6.2 (+3.9..+8.6)", "mde": null, "verdict": "adopted", "verdict_detail": "first no-harm composition gain", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2018" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-351", "code": "COMPOSE", "name": "COMPOSE", "family": "packs-delivery", "question": "Do 10-13 small packs compose?", "hypothesis": null, "preregistered": { "status": "yes (2d071770)", "where_public": null }, "set": "MAIN 92/97", "judge": "replay; Claude Haiku 204", "n": null, "seeds": null, "result": "delivered 22→29 (stock) / 23→32 (gidf), crowding flat; answers gidf +0.009 (−0.022..+0.039)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "delivery yes; answers flat", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2019" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-352", "code": "PRECISION-GATE", "name": "PRECISION-GATE", "family": null, "question": "Withhold wrong context (score gate / select gate)", "hypothesis": null, "preregistered": { "status": "yes (a350a88b)", "where_public": null }, "set": "MAIN 97", "judge": "Claude Haiku 38", "n": null, "seeds": null, "result": "P1 +0.016 (−0.047..+0.076) n 16; P2 +0.004", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "FAIL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2020" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-353", "code": "AUTOPSY52 / H4b", "name": "AUTOPSY52 / H4b", "family": "packs-delivery", "question": "Why 52 pack facts are not delivered; global avgdl", "hypothesis": null, "preregistered": { "status": "no / criterion", "where_public": null }, "set": "MAIN 92", "judge": "replay", "n": null, "seeds": null, "result": "SE packs lose on per-pack df cut, no chunk vectors, health misfire; gidf2 −1.4 vs gidf", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "H4b NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2021" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-354", "code": "F1b", "name": "F1b", "family": "packs-delivery", "question": "Chunk vectors for the cooking/money Q&A packs", "hypothesis": null, "preregistered": { "status": "yes (SEFIX 07ea2b5e)", "where_public": null }, "set": "MAIN 92/97", "judge": "replay; Claude Haiku 132", "n": null, "seeds": null, "result": "delivered 32→45 (article level; chunk level 32); answers −0.026 (−0.051..−0.002)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "delivers; answers FAIL", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2022" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-355", "code": "SEFIX-ALL", "name": "SEFIX-ALL", "family": "packs-delivery", "question": "Lexical fixes F1-F4 for Q&A packs + gidf", "hypothesis": null, "preregistered": { "status": "yes (07ea2b5e)", "where_public": null }, "set": "MAIN 97, 2 seeds; Opus CV5 42; Opus cm 22", "judge": "Claude Haiku; Claude Opus", "n": null, "seeds": null, "result": "chunk-level 28→47; pooled +0.023 (+0.001..+0.048); Opus CV5 −0.002 (−0.062..+0.061), CE halves; Opus where fact delivered (cm) +0.173 (+0.082..+0.266)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "LEVER-CANDIDATE, then withdrawn (F3 unsafe)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2023" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-356", "code": "F3-SAFETY / F3b", "name": "F3-SAFETY / F3b", "family": "safety", "question": "Is the F3 health no-source fallback safe?", "hypothesis": null, "preregistered": { "status": "yes (15be6d08", "where_public": null }, "set": "65 safety probes", "judge": "blinded agent reading", "n": null, "seeds": null, "result": "F3 fires 21/65, unsafe 9→13; F3b-A vacuous, F3b-B unsafe 9→12", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "FAIL; F3 withdrawn", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2024" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-357", "code": "SEFIX-124", "name": "SEFIX-124", "family": "packs-delivery", "question": "F1/F2/F4 without F3", "hypothesis": null, "preregistered": { "status": "yes (07ea2b5e)", "where_public": null }, "set": "MAIN 97, 2 seeds", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "chunk-level 42/92, 0 lost; pooled +0.013", "result_verified": true, "effect": null, "ci": "+0.013 (−0.008..+0.035)", "mde": null, "verdict": "adopted", "verdict_detail": "not a lever-candidate; adopted for delivery", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2025" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-358", "code": "SEARCHPAR R1-R5", "name": "SEARCHPAR R1-R5", "family": "packs-delivery", "question": "Dense re-rank, header vectors, BM25 k1/b, near-dup drop, sticky sources", "hypothesis": null, "preregistered": { "status": "yes (d8fca0ee)", "where_public": null }, "set": "MAIN 92; QFORM; F8 pairs", "judge": "replay", "n": null, "seeds": null, "result": "R1 mix +10 chunk-level facts; R2/R3/R4 NO; R5 subject carry +17/48 F8", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "R1 candidate; others NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2026" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-359", "code": "STACK", "name": "STACK", "family": "submitted-app", "question": "SEFIX + chunk vectors + R1 together", "hypothesis": null, "preregistered": { "status": "yes (c0fabf99)", "where_public": null }, "set": "MAIN 92", "judge": "replay", "n": null, "seeds": null, "result": "chunk-level 45 vs bar 52; routes are alternatives, not a stack", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2027" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-360", "code": "READER / COMBINED", "name": "READER / COMBINED", "family": "packs-delivery", "question": "Reader fixes (order, cite-or-ground, keep answers); SEFIX + fact-first", "hypothesis": null, "preregistered": { "status": "yes (f361f738", "where_public": null }, "set": "44 delivered ids; MAIN 97 × 2 seeds", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "none READS; 11/44 \"delivered\" had no fact in kept text; COMBINED +0.022 vs K3, fact-first hurts v0.1", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2028" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-361", "code": "CHUNKLEVEL", "name": "CHUNKLEVEL", "family": "packs-delivery", "question": "Re-score delivery at chunk level", "hypothesis": null, "preregistered": { "status": "n/a (0 LLM)", "where_public": null }, "set": "42 replays", "judge": "counts", "n": null, "seeds": null, "result": "F1b 45→32; SEFIX-ALL 55→47", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "correction", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2029" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-362", "code": "DETECTOR-HN / FT / FT2", "name": "DETECTOR-HN / FT / FT2", "family": "safety", "question": "Retrain the health detector against cooking/money misfires", "hypothesis": null, "preregistered": { "status": "yes (7b7bed00", "where_public": null }, "set": "held-out titles; probes; MAIN", "judge": "0 LLM reads", "n": null, "seeds": null, "result": "FP 3.4 %→0 %, but recall losses 76-185 positives and probe routes 47→13-40", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NOT ACCEPTED (all)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-22", "fact_ids": [ "F-2030" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-363", "code": "NIGHT-A-KIT", "name": "NIGHT-A-KIT", "family": "packs-delivery", "question": "Do more crypto packs add v0.1 facts?", "hypothesis": null, "preregistered": { "status": "yes (530025688)", "where_public": null }, "set": "v0.1 52; MAIN 92", "judge": "replay", "n": null, "seeds": null, "result": "28 = 28", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "not adopted", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2031" ], "chapter": "delivery", "source": "verify pass 2026-10-24)" }, { "id": "X-364", "code": "NIGHT-B-TABLE", "name": "NIGHT-B-TABLE", "family": null, "question": "On real items, are delivered-fact answers better than non-delivered?", "hypothesis": null, "preregistered": { "status": "yes (530025688)", "where_public": null }, "set": "MAIN 92 keyed, 2 seeds; Opus CV5 28", "judge": "Claude Haiku; Claude Opus", "n": null, "seeds": null, "result": "delivered 0.420 vs not 0.419; within-item arm−control +0.044 (+0.009..+0.082) n 42; Opus slice +0.013 (−0.044..+0.072)", "result_verified": true, "effect": null, "ci": "between items +0.001 (−0.071..+0.069)", "mde": null, "verdict": "closed", "verdict_detail": "NOT PUBLISHABLE", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2032" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-365", "code": "SUPPORT-FLAG 2x2", "name": "SUPPORT-FLAG 2x2", "family": "ruler", "question": "Same with a judged support flag", "hypothesis": null, "preregistered": { "status": "yes (93f342bd)", "where_public": null }, "set": "MAIN 97", "judge": "Claude Haiku 190", "n": null, "seeds": null, "result": "supported vs not 0.44 vs 0.43; within-item +0.040 (−0.008..+0.087)", "result_verified": true, "effect": null, "ci": "gap +0.005 pooled", "mde": null, "verdict": "closed", "verdict_detail": "claim \"delivered facts are answered well\" does not hold between items", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2033" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-366", "code": "W98", "name": "W98", "family": "safety", "question": "Cooking/money-source veto on the health route", "hypothesis": null, "preregistered": { "status": "no (traces)", "where_public": null }, "set": "65 probes; MAIN 97", "judge": "counts", "n": null, "seeds": null, "result": "silences 5/5 real health probes", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2034" ], "chapter": "closure", "source": "verify pass 2026-10-24)" }, { "id": "X-368", "code": "ENGFIX-1023 / CO-PTES / PLUGFIX", "name": "ENGFIX-1023 / CO-PTES / PLUGFIX", "family": null, "question": "Engine fixes (CO vehicles, digit runs, seed cue, PT/ES CO, plugin routes)", "hypothesis": null, "preregistered": { "status": "n/a (tests)", "where_public": null }, "set": "110 plugin rows; suites", "judge": "replay", "n": null, "seeds": null, "result": "digits 3.7 s→14 ms; route 89→103/110; suite 1,645 green", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "patches staged", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2036" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-369", "code": "TOPICS27 / SUCCINCT", "name": "TOPICS27 / SUCCINCT", "family": "packs-delivery", "question": "Are the app's 27 suggested topics answered with support?", "hypothesis": null, "preregistered": { "status": "yes (5f8ba772", "where_public": null }, "set": "27 topics with Vital pack (dev)", "judge": "Claude Haiku sentence audit", "n": null, "seeds": null, "result": "unsupported Detailed 0.468 vs Succinct 0.283; invented specifics 98 vs 7; KEEP 7→19/27", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "Succinct for 1.5-1.7B", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2037" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-370", "code": "ENG-BANNER-KEPT / UXT-135", "name": "ENG-BANNER-KEPT / UXT-135", "family": null, "question": "Why does the \"Not found\" line show when passages were kept?", "hypothesis": null, "preregistered": { "status": "n/a", "where_public": null }, "set": "TOPICS27; MAIN 97", "judge": "replay", "n": null, "seeds": null, "result": "false banner 37/97; patch switches 37/38 to \"passages found don't support\"", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "patch ready; recount closes", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2038" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-371", "code": "VITAL-MISSES / KEEPPIN / PIN2 / FMT1-SCAN", "name": "VITAL-MISSES / KEEPPIN / PIN2 / FMT1-SCAN", "family": "packs-delivery", "question": "Recover the 4 missed topics", "hypothesis": null, "preregistered": { "status": "yes (80316ed8", "where_public": null }, "set": "27 topics; MAIN 97", "judge": "replay; PIN2 Claude Haiku 24", "n": null, "seeds": null, "result": "4 losses at 4 stages; pins recover Kyoto but fail context clause; PIN2 screen thin +0.018 (−0.034..+0.068)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO / HOLD", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2039" ], "chapter": "small-engine", "source": "verify pass 2026-10-24)" }, { "id": "X-372", "code": "W78", "name": "W78 (BINARY / B / C / PACKFMT)", "family": "packs-delivery", "question": "Cut phone reading cost of the dense stage", "hypothesis": null, "preregistered": { "status": "yes (e9d5e44a", "where_public": null }, "set": "MAIN 97; QFORM 216", "judge": "replay", "n": null, "seeds": null, "result": "kit 125.09→21.26 MB per question; cold stage 1 620→92 ms; kept 96/97; facts 42/42; W78B NO", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "open", "verdict_detail": "ADOPT-CANDIDATE (phone unmeasured)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2040" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-373", "code": "PARA135", "name": "PARA135", "family": "prompting-style", "question": "1/3/5 paragraphs on the 1.7B", "hypothesis": null, "preregistered": { "status": "yes (e59b069a)", "where_public": null }, "set": "27 topics; v0.1 51", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "unsupported 0.33/0.50/0.52; U 0.228/0.233/0.218; CE 37/43/49 per 100", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "Succinct best; support breaks at paragraph 2", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2041" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-374", "code": "DEEPMODE", "name": "DEEPMODE", "family": "packs-delivery", "question": "Deep Research multipass on the 1.7B", "hypothesis": null, "preregistered": { "status": "yes (8753e6b7)", "where_public": null }, "set": "27 topics; v0.1 51", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "U 0.174 vs Detailed 0.224 vs Succinct 0.263; 2.6-5.6x slower", "result_verified": true, "effect": null, "ci": "vs Detailed −0.050 (−0.087..−0.015)", "mde": null, "verdict": "closed", "verdict_detail": "worse; hide", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2042" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-375", "code": "UXT-24", "name": "UXT-24", "family": "packs-delivery", "question": "Personal files above Vital; follow-up source keep", "hypothesis": null, "preregistered": { "status": "yes (c2faa6eb)", "where_public": null }, "set": "29 questions on the app's JVM harness", "judge": "replay", "n": null, "seeds": null, "result": "fact kept 10/14 (Vital on) vs 13/14 off; rescue 14/14", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "YES-WITH-GUARD", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2043" ], "chapter": "closure", "source": "verify pass 2026-10-24)" }, { "id": "X-376", "code": "STYLE-SEFIX", "name": "STYLE-SEFIX", "family": "packs-delivery", "question": "Does Detailed use delivered facts better?", "hypothesis": null, "preregistered": { "status": "yes (25478a97)", "where_public": null }, "set": "MAIN 97", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "dU −0.028; delivered −0.047; CE 25.8→46.4 per 100", "result_verified": true, "effect": null, "ci": "−0.028 (−0.048..−0.008)", "mde": null, "verdict": "adopted", "verdict_detail": "Succinct stays", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2044" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-377", "code": "PLUGBATCH-SELF", "name": "PLUGBATCH-SELF", "family": "models", "question": "Plugin-stress routes on our 1.7B engine", "hypothesis": null, "preregistered": { "status": "yes (8dfa7a8a)", "where_public": null }, "set": "110 public rows", "judge": "route replay", "n": null, "seeds": null, "result": "85/110; engfix 89; after PLUGFIX 103", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "measurement", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2045" ], "chapter": "small-engine", "source": "verify pass 2026-10-24" }, { "id": "X-379", "code": "U5-FOLLOWUP-PACKS", "name": "U5-FOLLOWUP-PACKS", "family": "packs-delivery", "question": "Does the turn-1 source survive the follow-up with the pack on?", "hypothesis": null, "preregistered": { "status": "yes (fab6359f)", "where_public": null }, "set": "48 F8 pairs; 20 long-session turns", "judge": "replay", "n": null, "seeds": null, "result": "survives 1/19 and 4/13; subject carry recovers 15/18 at 1.22x off-subject", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "no fix by the line", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2047" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-380", "code": "UXMATRIX", "name": "UXMATRIX", "family": "packs-delivery", "question": "In each mode, is the answer right when the pack holds the content, and what happens without it?", "hypothesis": null, "preregistered": { "status": "yes (0cc3ca20)", "where_public": null }, "set": "60 items (40 cm + 20 v0.1)", "judge": "Claude Haiku outcome labels", "n": null, "seeds": null, "result": "Succinct ON CWS 0.28 vs OFF 0.03; Detailed 0.18 vs 0.03; correct 35 vs 27 % (ON), 28 vs 17 % (OFF, approx.)", "result_verified": true, "effect": null, "ci": "ON−OFF CWS +0.27 (Succinct)", "mde": null, "verdict": "adopted", "verdict_detail": "Succinct beats Detailed", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2048" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-381", "code": "WARN-KEPT", "name": "WARN-KEPT", "family": null, "question": "Warn cited-but-unsupported answers", "hypothesis": null, "preregistered": { "status": "yes (d5b5436d)", "where_public": null }, "set": "UXMATRIX 120; MAIN", "judge": "0 LLM", "n": null, "seeds": null, "result": "warns 9/26 wrong at cost 5/37 correct", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2049" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-382", "code": "PACKPROOF", "name": "PACKPROOF", "family": "packs-delivery", "question": "With vs without the pack on its own self-tests", "hypothesis": null, "preregistered": { "status": "yes (37f74d6f)", "where_public": null }, "set": "180 self-tests of 9 packs (pack-written)", "judge": "Claude Haiku", "n": null, "seeds": null, "result": "correct-any 0.48 vs 0.26; when own article kept 0.65 vs 0.28; OFF wrong-confident 0.17", "result_verified": true, "effect": null, "ci": "+0.23 (+0.16..+0.31)", "mde": null, "verdict": "closed", "verdict_detail": "P1 MISS (bar 0.75); honest-when-not via warning", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2050" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-383", "code": "PACK-GAP", "name": "PACK-GAP", "family": "packs-delivery", "question": "Would new packs close the undelivered gap?", "hypothesis": null, "preregistered": { "status": "no (0 LLM)", "where_public": null }, "set": "50 undelivered MAIN items", "judge": "FTS probe", "n": null, "seeds": null, "result": "44/50 already in K13 (search problem); new packs ≤ ~1 fact each", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "no new packs", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2051" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-384", "code": "NOWIKI", "name": "NOWIKI", "family": "packs-delivery", "question": "Kit without simplewiki and Vital", "hypothesis": null, "preregistered": { "status": "yes (5d40edcd)", "where_public": null }, "set": "MAIN 97; 160 self-tests", "judge": "replay", "n": null, "seeds": null, "result": "MAIN delivered 42→33", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "KEEP general packs", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2052" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-385", "code": "QREWRITE", "name": "QREWRITE", "family": "packs-delivery", "question": "Ceiling of question rewriting (oracle)", "hypothesis": null, "preregistered": { "status": "yes (60fdc688)", "where_public": null }, "set": "MAIN 97; 180 self-tests", "judge": "replay", "n": null, "seeds": null, "result": "MAIN 42→51 (+9 < +10); ST kept 100→118 (+18 < +20)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO-CEILING (stopped at stage 0)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2053" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-386", "code": "PACKSHAPE / V2-CONFIRM", "name": "PACKSHAPE / V2-CONFIRM", "family": "packs-delivery", "question": "Factbook as short country-named sentences", "hypothesis": null, "preregistered": { "status": "yes (89f565ca", "where_public": null }, "set": "20 factbook self-tests (pack-written), 2 seeds", "judge": "Claude Haiku; Claude Opus", "n": null, "seeds": null, "result": "fact chunk kept 7→15/20; correct 5→10/20 (pooled 0.50); Opus U 0.316→0.436", "result_verified": true, "effect": null, "ci": "Opus +0.121 (+0.030..+0.221)", "mde": null, "verdict": "open", "verdict_detail": "LEVER-CANDIDATE; NOT-CONFIRMED by 1 new warned CE", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2054" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-387", "code": "ARTCAP / PACKSHAPE-FIRSTAID / D2Q", "name": "ARTCAP / PACKSHAPE-FIRSTAID / D2Q", "family": "packs-delivery", "question": "Article cap; shaped first-aid pack; doc2query questions", "hypothesis": null, "preregistered": { "status": "yes (4069ae83", "where_public": null }, "set": "MAIN 97; self-tests", "judge": "replay", "n": null, "seeds": null, "result": "MAIN 42→39-43; first-aid kept 5→4; d2q own-article 9→10", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO (all three)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2055" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-388", "code": "HEALTHKEEP-DIAG", "name": "HEALTHKEEP-DIAG", "family": "packs-delivery", "question": "Why does the health path keep zero chunks?", "hypothesis": null, "preregistered": { "status": "no (diagnosis)", "where_public": null }, "set": "40 health self-tests", "judge": "replay", "n": null, "seeds": null, "result": "topic-empty rule; fact chunk available in only 4 of 27 misses", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "diagnosis", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2056" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-389", "code": "X2B-BUDGET", "name": "X2B-BUDGET", "family": "packs-delivery", "question": "Wider append budget", "hypothesis": null, "preregistered": { "status": "yes (4ede4263)", "where_public": null }, "set": "MAIN 97; ST 180; probes", "judge": "replay", "n": null, "seeds": null, "result": "MAIN 42→47; but context 1.17x and 9 dose/number probes changed", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "NO", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2057" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-390", "code": "RETBENCH", "name": "RETBENCH", "family": "packs-delivery", "question": "0-LLM retrieval bench + 10-config sweep", "hypothesis": null, "preregistered": { "status": "yes (55386ee6)", "where_public": null }, "set": "MAIN 97; ST 180; topics; follow-ups", "judge": "replay", "n": null, "seeds": null, "result": "best k4+compress 0.4: MAIN 42→45, 0 lost; probes NO (2 dose/number)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "candidate, out of ENGINE-RC", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2058" ], "chapter": "delivery", "source": "verify pass 2026-10-24" }, { "id": "X-391", "code": "VALIDATION-1024", "name": "VALIDATION-1024 (0-LLM suites)", "family": "packs-delivery", "question": "Does ENGINE-RC pass identity, delivery, routes, cost, multi-turn suites?", "hypothesis": null, "preregistered": { "status": "yes (aa08650f + addenda)", "where_public": null }, "set": "MAIN 97; ST 180; 65 probes; 110 plugin rows", "judge": "replay", "n": null, "seeds": null, "result": "identity 97/97, 180/180; MAIN 42; ST fact 84 vs 77; routes 103/110; 21.26 vs 125.09 MB/q; V1 letter-FAIL pre-existing", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "PASS (V1 letter-FAIL)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-23", "fact_ids": [ "F-2059" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-392", "code": "VALIDATION judge", "name": "VALIDATION judge (styles)", "family": "ruler", "question": "Succinct vs S+ (one 4-5 sentence paragraph) vs two paragraphs", "hypothesis": null, "preregistered": { "status": "yes (addendum 5)", "where_public": null }, "set": "UXMATRIX 60", "judge": "Claude Haiku 478", "n": null, "seeds": null, "result": "correctness equal; wrong-unwarned ON 0.12→0.23/0.25", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "Succinct stays", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2060" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-393", "code": "VALIDATION judge", "name": "VALIDATION judge (follow-ups)", "family": "ruler", "question": "Does subject carry raise follow-up accuracy?", "hypothesis": null, "preregistered": { "status": "yes (df6c0c18)", "where_public": null }, "set": "48 F8 follow-ups", "judge": "Claude Haiku 118", "n": null, "seeds": null, "result": "correct 0.458 with and without carry; turn-1 source kept 0.08→0.71", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "UX gain, not accuracy", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2061" ], "chapter": "closure", "source": "verify pass 2026-10-24" }, { "id": "X-394", "code": "CLEAN-LITE", "name": "CLEAN-LITE", "family": "submitted-app", "question": "Apply the closed-app decisions to the clean native branch", "hypothesis": null, "preregistered": { "status": "n/a (build)", "where_public": null }, "set": null, "judge": "unit tests", "n": null, "seeds": null, "result": "Succinct only; 1.7B default; English only; debug APK built", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "tooling", "verdict_detail": "built; device tests not run", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2062" ], "chapter": "submission", "source": "verify pass 2026-10-24" }, { "id": "X-398", "code": "BOARLITE-HEALTHGATE", "name": "BoarLite health gate (B+FU detector, gate after retrieval)", "family": "safety", "question": "Can a lexical health detector in the native app reproduce ENGINE-RC's health gating?", "hypothesis": null, "preregistered": { "status": "no (build gate, 0 LLM; rules fitted on the probes)", "where_public": null }, "set": "65 F3-SAFETY probes (dev); 60 health self-tests; 251 + 40 non-health items", "judge": "none (0 LLM)", "n": "65 + 60 + 291", "seeds": null, "result": "41-42 of 65 probes fixed text incl. the 4 invented-dose answers (ENGINE-RC 41); 12 of 60 health self-tests fixed (Wikipedia-only articles); 0 new false fires on 251", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "shipped in the final candidate b4ef94f3", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2191", "F-2192", "F-2193", "F-2194" ], "chapter": "submission", "source": "update 2026-10-24 evening (QUEUE 1787-1815 + arm reports)" }, { "id": "X-399", "code": "SEARCHPORT P1", "name": "SEARCHPORT P1 (bench search fixes in the native engine)", "family": "packs-delivery", "question": "Does porting SEFIX-124 + head pass + X2b make the right chunk reach the 1.7B without health or safety changes?", "hypothesis": null, "preregistered": { "status": "yes (fa3b1144)", "where_public": null }, "set": "MAIN 92 keyed; 180 pack self-tests; 65 F3-SAFETY probes (dev)", "judge": "replay (0 LLM); regex count", "n": "92 + 180 + 65", "seeds": "1 (42)", "result": "MAIN 26 → 40, ST 38 → 64, context 1.07×; 3 new dose-figure probes", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "FAIL gate d (safety) after the one fix; switch off", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2195", "F-2196" ], "chapter": "submission", "source": "update 2026-10-24 evening (QUEUE 1787-1815 + arm reports)" }, { "id": "X-400", "code": "GROUNDED-DOSE A / SEARCHPORT Addendum 1", "name": "Search port re-gated under the grounded-dose rule", "family": "safety", "question": "With the widened health gate, does the port add ungrounded dose figures?", "hypothesis": null, "preregistered": { "status": "yes (GROUNDED-DOSE criterion; Addendum 1)", "where_public": null }, "set": "65 F3-SAFETY probes (dev)", "judge": "grounded-dose scorer (0 LLM)", "n": "30 changed probes", "seeds": "1 (42)", "result": "ungrounded 9 vs 5; 9 probes fixed → model (health passage shown)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "closed", "verdict_detail": "FAIL; NOT-WORTH for the submission", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2203" ], "chapter": "submission", "source": "update 2026-10-24 evening (QUEUE 1787-1815 + arm reports)" }, { "id": "X-401", "code": "GROUNDED-DOSE B", "name": "X2B-BUDGET 900 re-gated under the grounded-dose rule", "family": "safety", "question": "Does a wider append budget add ungrounded dose figures on the bench engine?", "hypothesis": null, "preregistered": { "status": "yes (GROUNDED-DOSE criterion)", "where_public": null }, "set": "65 F3-SAFETY probes (dev)", "judge": "grounded-dose scorer (0 LLM)", "n": "65 (12 changed)", "seeds": "1 (42)", "result": "ungrounded 7 vs 7; grounded 23 vs 23; 0 fixed → model", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "PASS by the letter; one grounded-yet-toxic misreading flagged; adoption not recorded", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2205" ], "chapter": "delivery", "source": "update 2026-10-24 evening (QUEUE 1787-1815 + arm reports)" }, { "id": "X-402", "code": "GROUNDED-DOSE C", "name": "RETBENCH B-k4comp40 re-gated under the grounded-dose rule", "family": "safety", "question": "Does the best retrieval config add ungrounded dose figures on the bench engine?", "hypothesis": null, "preregistered": { "status": "yes (GROUNDED-DOSE criterion)", "where_public": null }, "set": "65 F3-SAFETY probes (dev)", "judge": "grounded-dose scorer (0 LLM)", "n": "65 (3 changed)", "seeds": "1 (42)", "result": "ungrounded 7 vs 7; grounded 23 vs 23; 0 fixed → model", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "PASS; earlier rejection rested on a proxy", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2204" ], "chapter": "delivery", "source": "update 2026-10-24 evening (QUEUE 1787-1815 + arm reports)" }, { "id": "X-403", "code": "SEARCHPORT Addendum 2 (ON-NH)", "name": "Search port on except for health turns (ON_NO_HEALTH)", "family": "packs-delivery", "question": "Is the port worth it for non-health questions if health turns keep the old search?", "hypothesis": null, "preregistered": { "status": "yes (ADDENDUM-2, frozen 19:2x)", "where_public": null }, "set": "MAIN 92 keyed; 180 pack self-tests; 65 F3-SAFETY probes (dev); UXMATRIX 60", "judge": "replay; grounded-dose scorer; Claude Haiku labels (S2)", "n": "92 + 180 + 65 + 60", "seeds": "1 (42)", "result": "MAIN 26 → 40, ST 38 → 62, context 1.07×; S1 ungrounded 6 vs 4 (FAIL); S2 usefulness bar passed (CWS 9 → 15, correct-any 20 → 21, WU 31 → 30 of 51 changed UXMATRIX items; F-2208)", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "adopted", "verdict_detail": "frozen verdict not WORTH-IT (S1 FAIL by the letter); merged as ON_NO_HEALTH under the author's prior condition", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2207", "F-2209" ], "chapter": "submission", "source": "update 2026-10-24 evening (QUEUE 1787-1815 + arm reports)" }, { "id": "X-404", "code": "BOARLITE-SMOKE", "name": "BoarLite device smoke on a Pixel 8a (airplane mode)", "family": "device", "question": "Does the closed APK install, unpack, answer offline and stay off the network?", "hypothesis": null, "preregistered": { "status": "no (hand smoke)", "where_public": null }, "set": "device smoke (hand questions, Pixel 8a, airplane mode)", "judge": "none (hand check)", "n": "5 questions + candidate run", "seeds": "—", "result": "unpack 3 GB in 14 s; 0 B network; PSS 3.09 GB; candidate: model ready 18 s, first answer 11.9 s at 14.9 tok/s", "result_verified": true, "effect": null, "ci": null, "mde": null, "verdict": "measurement", "verdict_detail": "works offline; 5 findings (2 fixed in the candidate)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2198", "F-2201" ], "chapter": "phone", "source": "update 2026-10-24 evening (QUEUE 1787-1815 + arm reports)" }, { "id": "X-405", "code": "RULER-FINAL", "name": "Ruler v0.4 on BoarLite's final engine (desktop twin)", "family": "submitted-app", "question": "What is BoarLite's U-total on v0.4 under both judge families?", "hypothesis": null, "preregistered": { "status": "yes (ADDENDUM-RULER-FINAL, frozen 20:1x)", "where_public": null }, "set": "v0.4 (71 headline + 4 anchors)", "judge": "Claude Opus + OpenAI gpt-5.6-sol", "n": "71", "seeds": "1 (42)", "result": "U-total Claude 0.164 (0.123-0.208), OpenAI gpt-5.6-sol 0.245 (0.194-0.300); minimum over families 0.164 (0.123-0.206); win-rate 0 of 71 under both (1 gpt-5.6-sol tie)", "result_verified": true, "effect": "lower-family U-total 0.164, below the 0.50 bar (desktop twin of b4ef94f3, one seed)", "ci": "Claude 0.123-0.208; gpt-5.6-sol 0.194-0.300; minimum 0.123-0.206", "mde": null, "verdict": "measurement", "verdict_detail": "run complete; below the 0.50 bar on the lower judge family; published 2026-10-25 as history (desktop twin of b4ef94f3, before the Rat persona line); the headline is DEVICE-FINAL (OF-8)", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-24", "fact_ids": [ "F-2214", "F-2215", "F-2217" ], "chapter": "new-ruler", "source": "update 2026-10-24 evening (QUEUE 1787-1815 + arm reports)" }, { "id": "X-406", "code": "DEVICE-FINAL", "name": "DEVICE-FINAL: final on-phone matrix of LabRat (engine 7788924a; APK 0ad476ca)", "family": "device", "question": "What are LabRat's claim numbers measured on the phone (ruler v0.4 under both judge families; speed, memory, offline) and its product numbers (probes, pack self-tests, UXMATRIX)?", "hypothesis": null, "preregistered": { "status": "yes (criterion frozen 2026-10-25 03:3x, before generation)", "where_public": null }, "set": "bench v0.4 (exposed); 65 safety probes; 140 on-K13 self-tests; UXMATRIX 60", "judge": "Claude Opus + OpenAI gpt-5.6-sol (M1); Claude Haiku labels (M3, M4)", "n": "71 headline items + 4 anchors × 2 seeds; 65 probes; 140 self-tests; 60 UXMATRIX", "seeds": "42, 43 (M1; 44 dropped for time)", "result": "M1: win-rate 0 of 71 under both families; U-total Claude 0.165 (0.127-0.205), gpt-5.6-sol 0.242 (0.194-0.296), seeds 42+43; phone = desktop twin (paired +0.003 / -0.007, seed 42). M5: p50 13.9 s, first word 9.8 s, 11.8 tok/s, RSS 3.87 GB, 0 network. M2: 53 of 65 fixed text, 2 new dose figures. M3: correct 0.40 of 140. M4: correct 0.35 of 60", "result_verified": true, "effect": "lower-family U-total 0.165, below the 0.50 bar (on the phone, two seeds)", "ci": "Claude 0.127-0.205; gpt-5.6-sol 0.194-0.296", "mde": null, "verdict": "measurement", "verdict_detail": "complete; below the 0.50 bar on the lower judge family, 0 wins in 71 under both; interrupted once at 11:31 and resumed; seed 44 dropped for time", "reopen_if": null, "closed_row": null, "registry_public": null, "date": "2026-10-25", "fact_ids": [ "F-2227", "F-2228", "F-2229", "F-2225", "F-2238", "F-2239", "F-2240", "F-2241", "F-2242", "F-2243", "F-2244" ], "chapter": "submission", "source": "update 2026-10-25 final (QUEUE 1822-1833 + DEVICE-FINAL report and aggregates + public LabRat repo)" } ] }