[VerifiedOrganic]

What this round fixes

Precision round for the semantic continuity lane (still off by default). This PR:

• removes the control false positives on baseline (+2 quantity) and held-out-2 (+1 dead-character acting);
• brings the on-mode repair rate to <=10% per call (1.19-3.80% measured);
• fixes the shadow/on parity mismatch on feat(evals): held-out continuity contradiction corpus #72 by measuring parity from actual shadow and on passes;
• addresses the three AO review findings on d1925528.

Review findings

1. UTF-8 context windows (semantic_judge.rs): the ±240-byte dead-context window now rounds both edges to character boundaries; regressions cover multibyte characters crossing the left and right edges.
2. Judge-policy cache revision (semantic_judge.rs): SEMANTIC_JUDGE_POLICY_VERSION is now semantic-judge-policy-v3; cache-key construction is shared, and a test seeds a v2 full report (which still parses as a contradiction) and proves a v3 lookup misses it.
3. Shadow parity from actual results (semantic_eval.py): the evaluator runs a fresh on pass, then a real cache-only shadow pass, keeps each mode's own results, and computes parity from the two actual summaries. No derived shadow view is used as evidence.

General precision fixes

• Quantity contradictions require balance/stock evidence, not a price or transaction.
• Dead-character state uses only states strictly before the scene; memory/flashback/vision framing is rejected.
• Deterministic dedupe now covers same-subject/same-scene findings for families beyond world rules (this removed the baseline dead-character duplicates); secret dedupe also requires the same secret fact, and the remaining families require their own canon/event/fact identity, so an unrelated scene-level finding cannot suppress an independent conflict.
• Family evidence gates run on compact, full, and host reports before caching or emission, so the supported report forms cannot bypass the precision policy.
• Secret disclosures require a distinctive value detail or a public-audience marker, so a private accusation that only names the topic is not a leak.
• No set-specific rules; all gates are family/evidence/general.

Results

Judge ollama:gemma4:31b, temperature 0, seed 7; same config for all sets. on is the fresh dispatch pass; shadow reuses the validated-report cache.

Set	Deterministic	Semantic on	Warning precision	Control warnings	On repair	Shadow/on parity	Anchor validity	p95 prompt/output/latency
Baseline	28/28	28/28	n/a (0 warnings)	0	1.18% (85 calls)	pass	1.00	2,173 / 627 / 1,797 ms
#72	54/56	54/56	1.00	0	2.63% (76 calls)	pass	1.00	2,893 / 1,036 / 2,507 ms
Held-out-2	55/56, 56/56 type	55/56, 56/56 type	n/a (0 warnings)	0	0.00% (66 calls)	pass	1.00	2,029 / 646 / 1,734 ms
Held-out-3	26/56, 28/56 type	29/56, 30/56 type	1.00	0	4.17% (72 calls)	pass	1.00	3,216 / 951 / 2,466 ms

Control warnings do not rise on any set. Warning precision is 1.0 where warnings exist. Evidence-anchor validity is 1.0. On-mode repair is <=4.17% of calls. Shadow/on would-be finding counts match on every set (the cache-only shadow pass makes only 0-6 retry dispatches, so its own repair ratio is reported separately in the checked-in reports and is not the gate).

Gates

• Deterministic --semantic off unchanged: baseline 28/28 with 3 control findings; feat(evals): held-out continuity contradiction corpus #72 54/56 with 0; held-out-2 55/56 + 56/56 type with 0; held-out-3 26/56 + 28/56 type with 0 control warnings.
• No control-warning rise; warning precision >=0.85 (1.0 where defined); on repair <=10% per call; anchor validity >=98%; one repair per call.
• cargo fmt --all -- --check, cargo clippy -p spindle-core -p spindle-adapters --all-targets -- -D warnings, cargo test -p spindle-adapters --lib (900 passed), cargo test -p spindle-core --lib (358 passed).
• Default remains semantic_mode = off; activation gates still not met because aggregate semantic lift is +3 localized/+2 type-present (all held-out-3).
• Held-out-4 untouched; no sealed-corpus data was read.

Checked-in reports: evals/results/semantic_precision_*.json.

[VerifiedOrganic]

Changes requested. Required fixes: make death-context slicing UTF-8 safe, invalidate cached reports produced before the new policy filters, and measure shadow parity/cost from actual shadow results rather than a copy of on-mode output. Reviewed d192552 against ed1a1d5. Static review only; no tests or project programs were run and the checkout was not modified.

[VerifiedOrganic]

Changes requested. Two correctness issues remain: the new evidence gates are bypassed by full/host reports, and deterministic dedupe can suppress distinct canon conflicts. Static review against origin/main; no programs, builds, or tests were run.

[VerifiedOrganic]

Changes requested. The full/host evidence-gate fix and secret-specific identity check address the concrete cases from the previous review. The general dedupe fallback still suppresses distinct canon conflicts in other families; one required inline finding describes the remaining issue. Static review against origin/main, including the follow-up diff; no programs, builds, or tests were run.

[VerifiedOrganic]

Approved — ready. The previously reported issues are addressed: shared evidence gates cover compact, full, and host reports, and deterministic dedupe now requires canon identity for independent fact/event families while preserving character-scoped death dedupe. The added regressions cover unrelated subjectless timeline candidates and multiple knowledge facts for the same character. No further high-confidence findings requiring changes. Static review of the diff against origin/main and the follow-up changes at ff8626f; no programs, builds, or tests were run.