The diff is complete. It does not touch AGENTS.md, content/ records, citations or supplied material, so the question is only whether the packet change weakens the §3.15 panel check. It does not. `tierOf` returns `diffTier` unchanged for tiers 0–1, and a test confirms the governance surface and canon are never reordered. Every omitted file is still named, the protocol is untouched, and on the stated figures unique records (gate verdicts, spend rows, run records) move into view, while the newly cut assessments (2→7 on #418, disclosed through the counts) stay as protected digests in the run account. `replyDifferences` is conservative: it compares strings exactly at every depth and only against the same run's installed files (by `producedBy`). Replies that cannot be parsed or are not carried whole are promoted ahead of the rest of tier 2, and the moved `parseYamlReply` has the same logic, so verb and comparison read identically. The DECISIONS.md correction serves §3.13/§3.14: it quotes the old wording in place, restates ten measurements as ranges without predicting, and its arithmetic checks ($2.33/$0.51≈4.6, $2.33/$0.33≈7.1; 1+3+1+2+1+1=9 judgments; 7+2+5+13+2=29 files). One residual gap, an imperfection rather than a violation: "carried whole" compares parsed fields only. YAML comments, text after a closing fence (the test's "Sincerely, the model" counts as carried) and parser-dropped trailing lines go uncompared, and the carried note drops the file's size. The account's stated terms should say this and the note should keep the size, so the PR body's "a reply cannot carry text past the comparison" is not overread by future seats.

The change strengthens rather than bypasses panel scrutiny: raw replies are deprioritized only after a recursive comparison with an assessment from the same run, while dropped fields, altered strings, malformed replies, and unmatched runs prevent the reply from being marked carried. The packet and omitted-file notices disclose the comparison terms and identify the carrying assessment, and tests cover the principal false-equivalence cases. The cost generalization is corrected with calibrated language, concrete run and judgment locators, preserved prior wording, and a new dated decision entry, satisfying uncertainty and revision-history requirements. No AGENTS.md edit, content citation, confidential material, fabricated locator, or unratified assessment is introduced.

This pull request directly addresses review note #420 by calibrating DECISIONS.md to report exact empirical measurements of model execution costs rather than projecting broad generalizations (§3.9, §3.13). It also enhances the panel's packet construction machinery in `arbiter-core.mjs` and `scripts/arbiter.mjs` by mechanically comparing raw check seat replies against installed assessments: raw replies that are provably carried whole are deprioritized when capping diffs so unique governance and spend records are preserved in the diff. The mechanical verification of 'carried whole' is strict, tested extensively against schema mismatches, string alterations, and stamp lengths, without weakening any constitutional check or editing AGENTS.md.

The diff touches only arbiter packet assembly, a shared YAML-reply parser, check.ts re-export, vendors comment cost, tests, and DECISIONS.md—no AGENTS.md edit, no content citations, no confidence material, and no run records to mis-stamp. Part 1 strengthens §3.15 visibility: runAccount now field-compares each raw seat reply to assessments that run installed (stamps bounded, one parser), returns carried paths, and capDiff prioritizes governance/inbox and uncarried replies while cutting only replies proven carried whole, with omittedNotes labeling that fact. Part 2 answers #420 by correcting the over-generalized cost sentence in place and restating measured check vs Arbiter figures without predicting future packets (§3.9/§3.13). Tests cover non-carried cases; nothing loosens mechanical or panel checks.

This is tooling and record-keeping work, not content, and nothing in it edits AGENTS.md, fabricates a citation, or exposes confidence material. The packet-reordering change strengthens rather than weakens §3.15's checks: it responds directly to the #418 parking (seats said 'unsure' because gate records, spend rows and run records were cut) by putting the files that exist nowhere else first, and the new 'carried whole' shortcut for raw seat replies is guarded by an exact, every-field, every-depth mechanical comparison whose terms are stated in the packet itself, whose six stamp exceptions are capped at 120-char scalars so text cannot be smuggled past the comparison, and whose tests cover precisely the cases that must NOT count as carried (changed string, dropped field, added field, wrong-run assessment). The capacity limitation is disclosed honestly — files still cut when a sitting overflows the packet are each named. The #420 correction models the right behavior under §3.14/§3.13: the overgeneralized sentence ('five to seven times... on a check') is corrected in place with the prior wording preserved in the note, the new entry distinguishes the one completed measurement from the nine Arbiter judgments and explicitly declines to predict future packets, and the internal arithmetic checks out ($2.33/$0.51 ≈ 4.6, $2.33/$0.33 ≈ 7.1, matching the corrected range). The protocol file the seats judge under is untouched, and the effort setting is left as the founder's decision. I see no rule broken; the unverified historical cost figures are internal ledger records the DECISIONS entry itself says how to reconstruct, not external citations.