The diff does not edit AGENTS.md, touches nothing under content/, and adds no citations (the tooling found none). It publishes no confidence material: the change is docs, a descriptive count under proposals/, reader wording, a scheduler fix and a budget fix. The check.ts change tightens a check rather than loosening it. Production still uses assertSeatsWithinBudget when no deps are injected, the repair call now goes through the same guard with outstanding calls counted, a refused seat is recorded failed with the cap named, and a test pins that order. busy-cases.mjs is a scheduling courtesy, not a §3.15 check, and the PR discloses that it has no test coverage. MAINTENANCE.md now states the gate rule as the amended §3.15 has it, and the arbiter code is untouched, so nothing is reclassified or weakened. The DECISIONS entries are candid: they correct #445's uncomputed claim with arithmetic that checks (3 of 5 vs 2 of 4 within one step on GEO-C030) and keep #452 parked. They also record the operator's reading of the Engelbach Gutenberg text while explicitly declining to treat it as what the ledger holds (§3.8/§3.9); I see no internal contradiction there and do not treat unfamiliarity as evidence. The split-count README is internally consistent: 126/2, 34/20 and 14 only-higher / 6 both-ways all recompute from its own table under the §3.15 scale. It calls itself a description, has a 'what it does not show' section, and changes no protocol on its strength (§3.12, §3.13). Two small items deserve an answer, not a park: the standing.ts header adds an unmentioned clarification (a pre-stamp reconsideration carrying `reconciles` is read by that stamp alone) that should be confirmed against the code and the 2026-08-26 decision, since it sits in tension with AUTOMATION.md's 'a stamp … never take from it'. And LatestStrip relabels 'last content update' as 'the change history, last entry', which is accurate only if lastContentUpdate is the last displayed history entry (§7).

The added finding in AUTOMATION, DECISIONS, and the experiment README says that on 14 of 20 split claims “every seat that differs” grades higher. The report’s own AMZ-C023 row contradicts that literal claim: the judgment is `unresolved`, while one seat says `weakly_supported`, which is lower under both the constitutional ordering and the script’s rank, while other seats grade it higher. `count.mts` reveals that direction is actually classified only among `far` verdicts—those more than one step away—not among every verdict whose word differs. This mislabels a quantified observation and presents a falsely exact directional finding; reword every instance to “every seat more than one step away” if that is intended, or recompute the count over all differing verdicts.

This pull request touches no case content under content/ and makes no amendments to AGENTS.md. It updates documentation (docs/AUTOMATION.md, docs/DECISIONS.md, docs/MAINTENANCE.md) to accurately record the project's state as of 2026-09-30, including a formal correction regarding #445 and documentation of the 2026-09-09 founder amendments. The code changes in src/pipeline/check.ts strengthen budget enforcement during repair rounds with full test coverage, and scripts/busy-cases.mjs appropriately extends the busy-case detection to all open pull requests. User-facing strings in the UI and method page clarify independence and noise without weakening any constitutional check.

The diff updates status docs, records operational findings/corrections in DECISIONS, and mends two loop faults (budget re-check before a check repair; busy-cases reading every open PR branch). No AGENTS.md edit, no case ledger/content changes, no run records, and citation tooling found no new DOI/arXiv/URL identifiers. The standing.ts header and MAINTENANCE gate wording are aligned to the existing 2026-09-09 amendment rather than loosened; reader-facing “case rests on” copy leaves the ledger term intact. The panel-split count is an offline description over already-published checks with an explicit non-adoption of protocol change. Nothing in the material positively shows fabrication, confidence exposure, provenance mis-stamp, or check-weakening.

No AGENTS.md edit appears in the diff, so the constitution rule is not triggered. The MAINTENANCE.md lane rewrite does not weaken the gate — it restates the pass/park rule exactly as the founder amendment of 2026-09-09 already words it in §3.15 (one seat parks alone only for fabrication, confidence material, or a constitution edit; two seats otherwise; lone other objections become review notes), and the check.ts change strengthens enforcement by putting the repair call under the same budget guard, with a test pinning the order. No citations are added (verification found none), no case content is touched, and the decisions entry is a model of §3.14-style honesty: it records that the operator's earlier #445 claim was asserted without computation and was false, that the panel stays parked on GEO-E002 because the record's statement of its source outruns its own stated limitation (§3.9), and that the case reads 'not yet ratified' — presenting no unratified material as ratified. The panel-splits finding is explicitly labeled a description with nothing run for it, its script and its limits are shipped beside it, and no protocol or standing rule is changed on its strength. Two items I cannot verify from the packet but which do not rise to violation: the intake.ts comment asserting inputHash is 'null on every check run before' 2026-09-30 describes historical records I cannot inspect, and the LatestStrip relabel keeps the same value and changelog link — if the maintainer wants that tightended, the label should be shown to track actual content updates rather than any changelog entry, but that is a display nuance under §7, not an epistemic breach.