[VerifiedOrganic]

Diagnosis (public sets only; held-out-4 untouched)

Measured with the pre-fix judge (ollama:gemma4:31b) on #72 and held-out-3. Reports: evals/results/semantic_p4_held-out-72.json, evals/results/semantic_p4_held-out-3.json.

Set	Judgments	Contradiction	No-contradiction	Uncertain	Confidence >=0.85	Discards	Stop reasons	Repair	p95 output
#72 shadow	200	0	162	38	200	not_contradiction 200	success 45, truncated 52	34.0%	1,200 (2,744 chars)
held-out-3 shadow	119	0	100	19	119	not_contradiction 119	success 31, truncated 75	39.6%	1,200 (2,789 chars)

One sentence: zero findings reached even info because the judge never returned a contradiction verdict; every validated judgment was compatible or insufficient at high confidence, so policy had nothing to promote, downgrade, or discard.

No evidence-anchor, canon-ref, threshold, or dedupe discards were recorded. The dominant failure was output shape/length: 52/97 #72 calls and 75/106 held-out-3 calls truncated at the 1,200-token cap while emitting the verbose report contract.

Root causes and fixes

• Verbose output contract → compact per-candidate wire format: one short record per candidate, contradiction-only conflict and evidence_quote; the server expands it into the existing validated report contract. Truncation fell to zero on the measured sets.
• Evidence offsets from the model → server resolves the quote span, tolerating whitespace normalization; exact spans are still verified. Evidence-anchor validity is 1.0.
• Overclaiming → general family rules plus a second, contradiction-only verifier call. The verifier rejects different quantity events, fact mentions as knowledge, non-disclosures, and action-only world-rule claims; unavailable verification fails closed.
• Duplicate claims → evidence-identity dedupe (scene + check type + byte span) in addition to structured P1 dedupe.
• Temporal/world-rule/secret evidence quality → contradiction requires a temporal quantity, rule-constraint overlap, and explicit disclosure/cost evidence respectively.

Stronger model check (one model, same sample)

ollama:deepseek-v4.1-flash was tried on the same #72/held-out-3 sample. It was not an improvement: 111/120 #72 calls and 121/125 held-out-3 calls truncated at the 1,200-token cap, with zero findings and ~8-9 s p95 latency. It is not being adopted; the operator still chooses the judge model.

Results (same head, gemma4:31b)

Set	Deterministic	Semantic on	Warning precision	Control warnings	Repair shadow/on	p95 prompt/output/latency (shadow)
Baseline	28/28	28/28	0.25	+2 quantity	1.1% / 50.0%	2,209 / 626 / 1,831 ms
#72	54/56	54/56	1.00	0	2.6% / 33.3%	2,850 / 981 / 2,768 ms
Held-out-2	55/56, 56/56 type	55/56, 56/56	0.00	+1 dead acting	1.5% / 50.0%	2,038 / 663 / 1,886 ms
Held-out-3	26/56, 28/56 type	29/56, 30/56	1.00	0	4.2% / 50.0%	3,138 / 951 / 2,671 ms

Held-out-3 gains +3 localized/+2 type-present with 1.0 precision, no control rise, evidence validity 1.0, and 4.2% shadow repair. #72's warnings are true positives but duplicate already-detected plants, so they add no recall. Baseline and held-out-2 still emit false-positive control warnings. Shadow/on would-be counts match on baseline and held-out-3 but differ by one on #72 (8 vs 9). The four-set lift/control gates do not pass on this head; the lane stays off by default. Held-out-4 remains sealed.

Gates

• Deterministic results with --semantic off unchanged: baseline 28/28 (3 controls), feat(evals): held-out continuity contradiction corpus #72 54/56 (0), held-out-2 55/56 + 56/56 type (0), held-out-3 26/56 + 28/56 type (0 control warnings).
• Per-call prompt/output/latency and shadow repair ceilings pass.
• Warning precision/control gates fail on baseline and held-out-2; semantic lift is absent on baseline, feat(evals): held-out continuity contradiction corpus #72, and held-out-2.
• Full workspace tests, clippy -D warnings, fmt: pass.
• CI/review pending on the reviewed head.
• Default remains off; checkpoint posted; wait for maintainer merge.

Gates

• Deterministic results with --semantic off unchanged: baseline 28/28 (3 controls), feat(evals): held-out continuity contradiction corpus #72 54/56 (0), held-out-2 55/56 + 56/56 type (0), held-out-3 26/56 + 28/56 type (0 control warnings).
• Full workspace tests, clippy -D warnings, fmt: pass.
• CI/review pending on this PR.
• Default remains off; checkpoint posted; wait for maintainer merge.

Gates

PENDING final measurements; CI/review pending. The default remains off.

[VerifiedOrganic]

Changes requested. Three correctness issues need fixes: Unicode evidence anchoring, invalidation of pre-verifier cache entries, and deduplication of distinct canon violations sharing evidence. Reviewed head 8444ebf against base c3c5915. Static review only; no project programs or tests were run and the checkout was not modified.

[VerifiedOrganic]

Approved; ready. Reviewed head 1c32dd0 against base c3c5915 and checked the changes since the previous review. All three findings are addressed: normalized evidence offsets now map bytes correctly, cache identity includes a judge-policy revision, and shared-evidence deduplication includes canon and subject identity. Regression tests cover Unicode evidence and distinct canon references sharing a quote. No further required findings. Static review only; tests were not run and the checkout was not modified.

[VerifiedOrganic]

Approved; ready. Reviewed head 020b114 against base c3c5915, including the measurement/documentation changes since the previously approved revision. Implementation and fixes are unchanged. The refreshed reports support the documented recall/precision results and the decision to keep the lane off by default; no required findings.

Optional documentation cleanup: remove the duplicated section 4.3 heading, and clarify the held-out-2 parity sentence: its would-be warning totals are 1 vs 1, so any claimed mismatch must refer to warning identity rather than counts.

Static review only; no tests or project programs were run and the checkout was not modified.