The change touches only tooling (src/lib/arbiter-core.mjs), its tests, and a DECISIONS.md entry. It does not edit AGENTS.md, does not touch content/ or citations, and does not alter any vote threshold, parking rule, or CI check. It does not weaken the §3.15 concurrence check. It only reorders diff-capping so that a drafter reply is cut first when the tool has found it reproduced by the edition and draft its run installed (strings exact, with two stated allowances), or, for a first reply, when everything it holds beyond the repaired reply is printed in the account. A reply with any difference stays in the diff at its old OVERLAY_TIER, and the omitted note says where the reply was compared. Compared with the #431 situation, where a reply could be cut with no accounting at all, this gives the panel more information, and the one uncompared field (researchStatus) is disclosed on each reply's line. The DECISIONS stamp "direction in session, 2026-09-29" follows the convention the preceding entry explains: it gives the date of the direction, and the founder has not reviewed these changes. I see no provenance misattribution. Imperfections the operator should fix: (1) the component projection to {label, state, note} silently drops any other key in a reply's components, which contradicts the stated rule that a field the verb does not install is a difference; (2) duplicate claimIds are checked only through the first `find` match; (3) a non-string question or a non-array accounts value is skipped rather than flagged; (4) a reply with non-empty researchStatus is still cut first even though that field is uncompared. The fix would be to flag extra component keys, duplicates, and malformed fields, and either compare researchStatus against the research file or keep such replies out of CARRIED_TIER.

The new omission path can mark an edition reply “carried whole” and give it CARRIED_TIER even though `researchStatus` is expressly excluded from `editionReplyDifferences`. The account prints only the number of those entries and does not compare their values with the research files said to receive them. Consequently, the raw reply may be omitted while retaining unique text or provenance details that the panel has neither seen nor mechanically traced into installed records. That replaces scrutiny of the file with an incomplete equivalence check and therefore weakens or dodges the §3.15 gate. Compare every `researchStatus` entry against the exact installed research records, or keep any reply with nonempty `researchStatus` in the diff.

This pull request updates the panel packet generation tooling (`src/lib/arbiter-core.mjs`) to mechanically verify and summarize drafter replies against installed edition and assessment files, mirroring the existing mechanism for check seat replies. AGENTS.md is not modified, no content or epistemic checks are weakened, and all added logic is thoroughly covered by tests. The change enhances packet transparency and inspection efficiency without compromising constitutional gates.

The diff only extends run-account tooling in arbiter-core.mjs (plus tests and a DECISIONS.md log entry) so edition drafter replies are mechanically compared to the installed edition and draft assessment, allowing carried replies to be omitted with an explicit note. No AGENTS.md edit, no content/ ledger changes, no citations, and no confidence material. The comparison strengthens panel visibility rather than loosening §3.15 checks, and the decision entry is stamped to the founder direction already on record.

The change touches only docs/DECISIONS.md (the operating record, not the reserved constitution file) and tooling/tests; no content/ records change and no citation identifiers are added, so no fabrication question arises. Far from weakening §3.15's checks, it extends the #422 seat-reply comparison to edition runs: the repository's own tooling now mechanically verifies field-by-field that the verb installed exactly what the drafter wrote (strings exact, with only two narrow, stated allowances — the verb's 'Measured by the verb' rationale paragraph and the one-character bracket/brace span closure), and it prints a first reply's own divergent text so a seat need not say 'unsure' for unseen files as happened on #431. Cutting a carried reply from the packet does not hide material from the panel, because what 'carried' means is that the reply's text is present in the installed edition and draft assessment, which the account and diff display directly; researchStatus, though uncompared, lands in the research file that the records digest covers, and a reply the comparison faults stays in the diff. A first reply past the 12,000-character print limit is not marked carried and is left to the diff, preserving scrutiny rather than dodging it. The terms of the comparison are stated in the account and in the updated header, so the panel knows exactly what 'carried whole' rests on — honest labeling consistent with §3.8 and §4. I note one minor robustness issue for the operator's record: in omittedNotes, an edition reply omitted from an over-cap diff without being carried could render an undefined comparison target in the omitted-list note; that is routine polish, not a constitutional violation.