The core change in src/domain/judgment.ts replaces the sorted-verdict comparison with matching by label. Two components that trade verdicts are now named as two moves. Re-worded or re-cut components are printed on both sides with nothing claimed about them, and the 'no … component verdict … changed' sentence can be written only when every label and verdict was kept. That tightens the §3.14 change record rather than loosening it, and the new tests cover trades, re-wording, partial matches, empty lists, repeated labels and reordering. The public-facing edits move toward accuracy: the method page, AUTOMATION.md and the ComparisonPanel comment stop claiming full blindness and disclose that the evidence marking identifies the newer telling when the ledger has moved. Likewise, comparison.ts no longer asserts 'because the evidence had changed' for an undecided comparison, where the edition may have gone out to answer the panel. The readerBudget.ts comment reconciles two word counts and says which counter produced the kept pair, with no internal inconsistency. Nothing under content/, no protocol text and no AGENTS.md is touched, and the citation tool found no identifiers. DECISIONS.md is appended rather than rewritten, so the superseded 'a relabelled component is not' wording stays on the record. The two pre-fix editions are not published in this diff; the change that carries them must include the stated measurement and will be judged on its own. The Supervised-by stamp reuses the 2026-09-29 session date for a 2026-09-30 entry, which fits a continuing session and is not positive evidence of misattribution.

The change strengthens change-history accuracy by matching component verdicts by stable labels and explicitly reporting unmatched before/after components rather than inferring that nothing changed. Its tests cover verdict trades, relabeling, duplicate labels, additions, removals, and reordering, reducing the risk of misleading rationale records under §3.14. The decisions-log addition records the correction append-only and stamps the claimed founder supervision with date and channel, consistent with the provenance requirements. It neither edits AGENTS.md nor weakens ratification, citation, or provenance checks; the comparison-language corrections instead disclose when evidence labels reveal which edition is newer.

The change directly resolves review note #427 by fixing `src/domain/judgment.ts` so that component verdict changes are tracked by component label rather than a sorted list of states. This prevents swapped verdicts from silently collapsing into a false claim that no component verdict changed, upholding §3.14 (versioning and accurate change recording). The explanatory updates to `app/method/page.tsx`, `docs/AUTOMATION.md`, and `src/domain/comparison.ts` accurately align documentation with the actual blind comparison protocol. No content records or constitutional files were modified.

The diff fixes judgment change-accounting so components are matched by label, stopping false “no component verdict changed” reports when labels trade or are re-cut — directly strengthening §3.14 rather than weakening checks. Method/AUTOMATION/ComparisonPanel text is aligned with the existing compare protocol (evidence marks when the ledger moved; third publish path named), and the undecided-comparison sentence no longer overclaims “evidence had changed.” Word-count comment in readerBudget.ts is corrected to the markup-stripped figures the budget uses, with the counting method stated. No content/, citations, run records, confidence material, or AGENTS.md edits; DECISIONS.md records the review-note adoption with clear provenance.

This is a code and documentation fix with no content/ changes, no AGENTS.md edit, and no citations added (verification found none). The core fix in judgment.ts makes the change-detection verb stricter, not weaker — a verdict trade between components now yields two named changes instead of a false 'no component verdict changed' — which serves §3.14's requirement that revision histories record what changed rather than silently under-report it. The doc corrections (method page, AUTOMATION.md, ComparisonPanel comment) align descriptions with what the comparison protocol actually does (tellings marked with their evidence, which marks the newer when the ledger has moved) — replacing an inaccurate 'fully blind' description with a more honest one, the opposite of check-weakening; §3.9 is also served since what the verb cannot match is now printed and explicitly not claimed. The readerBudget.ts comment now gives the correct markup-stripped counts and states how they were counted, resolving a numeric discrepancy honestly rather than hiding it. The DECISIONS.md entry is dated, names its origin (review note #427, seat objections on #425) and carries the same founder-supervision stamp format as the preceding entry; I cannot verify the 2026-09-29 session from the packet, but per temporal humility unfamiliarity is not evidence of fabrication, and nothing in the material contradicts itself.