This change only tightens when the run account may call an edition reply 'carried' and cut it from the diff. researchStatus entries, which were counted but not compared, must now be found on the named research item with matching status and statusNote. A missing item, an unreadable research file (the default when no file is passed, since research = null) or a non-list entry is each a difference, so every new branch fails closed toward showing seats more, and no existing comparison is removed. The three further checks (component keys beyond label/state/note, a claim assessed twice, a question/accounts/article/rationale of the wrong kind) are also purely additive, and the new tests cover each branch, including the account keeping an untraced reply in the diff with carried left empty. Nothing under content/, no citation, no protocol text and no AGENTS.md changes, so the constitution, confidence and check-weakening concerns of §3.15 do not arise. The DECISIONS entry describes the change accurately and credits the OpenAI seat's #438 and the Anthropic seat's notes. Its 'Supervised-by: founder (direction in session, 2026-09-29)' stamp repeats the stamp on the #437 entry it amends, and I see no positive evidence of misattribution. One residual gap for a follow-up, not a violation: an accepted researchStatus entry is checked only for status and note, so an entry with another key (e.g. a 'reason' holding prose) would still be dropped unread while the reply is called 'carried whole', contrary to the doc comment's 'at any depth'; flag any entry key other than id/status/note, as is now done for components. Relatedly, accounts entries are String()-coerced rather than type-checked, and the carried line's 'status and note as written' overstates for no-op open/empty-note entries on items that already hold a note.

The revised check can still be dodged because `editionReplyDifferences` does not reject extra fields inside a `researchStatus` entry. For example, `{id:"X-R001", status:"answered", note:"Study X-S001 collected the rows.", aside:"unexamined text"}` matches the research item and produces no difference, so `runAccount` calls the reply “carried whole” and may remove it from the diff even though `aside` was neither installed nor reported. That contradicts the new comment and account language claiming every field is compared, and defeats §3.15’s fail-closed, inspectable review process. The entry must be required to be a mapping with only the installed fields—or every additional nonempty field must be reported as a difference—before it can be called carried whole.

The change directly addresses a previously identified gap where `researchStatus` entries in working reply files were counted rather than verified against the target research item, preventing dropped or refused notes from being silently omitted from review diffs. In `src/lib/arbiter-core.mjs`, `editionReplyDifferences` now verifies that each `researchStatus` entry matches the status and note recorded in the case's `research.yaml`, and adds stricter validation for extra component fields, malformed types, and duplicate claim assessments. These changes strengthen fail-closed mechanical verification under §3.15, leave AGENTS.md untouched, and add thorough unit tests.

The diff tightens run-account comparison of edition replies: researchStatus entries are now traced to the head-revision research file (status and note as written), and the three gaps named on #437 (extra component fields, duplicate claim assessments, wrong-kind fields) become differences instead of silent pass-through. That strengthens §3.15 inspectability rather than weakening it. docs/DECISIONS.md only appends a dated decision entry for review note #438; AGENTS.md is untouched; no content/, citations, confidence material, or run records are added. Tests cover carried, refused, missing-item, unreadable-file, and wrong-kind cases. Complete diff with nothing omitted.

This change tightens the repository's own verification tooling rather than weakening it: `editionReplyDifferences` previously counted `researchStatus` entries without comparing them, so a refused entry's text could sit in a reply that was nonetheless called carried — exactly the gap review note #438 named. The fix traces each entry to the research item it names, adds differences for untraceable entries, unreadable research files, duplicate claim assessments, wrong-kind fields, and extra component fields, and fails toward keeping the reply in the diff (fail-closed direction, consistent with §3.15's mechanical checks). The new tests in arbiter.test.ts cover each difference case, including the account-level carry/refuse behavior. No content/ files change, no citations are added (verification confirms none), no AGENTS.md text is touched — the DECISIONS.md entry is the operator's decisions log with the founder's supervision stamp, which §3.15 permits. I see no fabricated provenance, no confidence material, and no check dodged; the lone reservation I'd note for the record is that the comparison runs against the head-revision research file rather than the revision the run installed, but that only risks spurious differences (keeping replies in the diff), never silently passing text through.