The diff changes only tooling (src/lib/yaml-reply.mjs, src/lib/arbiter-core.mjs), a test file and an appended DECISIONS.md entry. It does not edit AGENTS.md, adds nothing under content/, adds no citations, and omits no files. For the §3.15 review machinery it tightens rather than loosens. readYamlReply yields the same parsed value as the old parseYamlReply, since the fence stripping and the drop-up-to-5-lines loop are unchanged. It also surfaces the fence, trailing text and parser-attached comments, so a reply with more than 200 uncompared characters can no longer be marked 'carried whole' and cut first from a seat's packet; below that, the count is printed on the account line. The account's terms now say what is not compared, and the omitted-file note gets back the file's size, so seats see more, not less. The tests match the code: 7+1+205=213 outside characters for the fenced paragraph, 202 comment characters for '# ' plus 201 x's, and exactly 200 stays carried. The DECISIONS entry openly corrects #422's overstatement ('a reply cannot carry text past the comparison'). Its Supervised-by line is dated, names its channel, matches the preceding entry, and shows no positive sign of misattribution. Residual gaps are worth a follow-up but are not violations: up to 200 uncompared characters, enough for a one-line retraction, can still sit unseen in an omitted 'carried' reply, and since the allowance is small the account could simply quote that text inline. Comments the parser does not attach, and directive text, are neither compared nor counted, so the terms' 'the YAML's own comments are … counted' slightly overclaims.

The change strengthens rather than weakens the assessment check: previously unlimited text outside parsed YAML could be ignored, whereas replies with more than 200 such characters are now refused the “carried whole” treatment. Within the allowance, the run account expressly reports the uncompared character count and defines the comparison’s limits, so the provenance claim is qualified rather than silently overstated. Omitted-reply notes also regain file size information, improving inspectability, and the added tests cover trailing text, YAML comments, the allowance boundary, explanatory terms, and omission metadata. The decision entry records the change, rationale, tests, and supervision with date and channel, consistent with versioned operational provenance.

This pull request tightens the arbiter's audit tooling by accounting for unparsed text and comments when checking whether raw seat replies are carried whole by installed assessments. Rather than weakening checks, it closes an epistemic blind spot where unparsed characters or YAML comments could otherwise escape comparison. The implementation introduces strict limits (200 characters max for outside text/comments) before disqualifying a reply from being considered carried, restores file size indicators to omitted notes, and adds thorough test coverage without altering the constitution.

The diff tightens run-account honesty for the #422 'carried whole' comparison: readYamlReply now returns set-aside outside text and YAML comments; uncompared length is counted and disclosed; over REPLY_UNCOMPARED_CHARS (200) a reply is not treated as carried; omitted-reply notes regain file size; section terms state what is not compared. That closes a disclosure gap rather than loosening §3.15 checks. No AGENTS.md edit, no content/ citations or confidence material, no fabricated provenance, and DECISIONS.md records the founder-directed fix. Tests cover the new bounds. Panel protocol is unchanged.

This change strengthens, not weakens, the §3.15 checks: previously text outside a seat reply's parsed fields (fences, dropped trailing lines, YAML comments) was silently invisible to the carried-whole comparison; now it is counted, bounded at 200 characters, stated on every carried line, and explained in the account's terms, so the hidden-text gap the Anthropic seat named on #422 is closed and disclosed rather than dodged. The new allowance is not a loosening — before this change an unbounded amount of outside text was ignored; now any excess over the documented allowance disqualifies the reply from being counted as carried, and the tests verify exactly that behavior. No AGENTS.md edit, no citation or content change, no run records, no provenance misstatement; the DECISIONS.md entry accurately records the change, its supervision, and its tests. The 200-character allowance for a fence and sign-off is a documented editorial threshold whose terms are displayed, which the constitution's transparency demands are met; ordinary parameter choice is not a violation.