The change is a presentation/infrastructure pass: `/panel` becomes `/operations` with links re-pointed, readers (config, spend read side, budget caps, the scheduler's pure choice) move from `src/pipeline/` into `src/domain/` as re-exports with no logic change I can see, and the page gains state, schedule, spend-against-caps, recent sittings and the review-note queue. AGENTS.md is untouched; no content/ records or citations are added (the verifier found none), and no §3.15 check is loosened — the ratification derivations in `standing.ts`/`panel.ts` are only re-labelled, and the harvest is additive. The new `governance/arbiter/*.yaml` and `governance/review-notes/*.yaml` are machine harvests stamped with `harvestedAt`, `judgedAgainst`, `promptVersion`, per-seat votes and costs; notably they display parked-but-merged verdicts (pr-206, pr-208, pr-212, pr-216) with every dissenting seat's words intact, which shows disagreement rather than hiding it, and `answerFrom` refuses closure-as-answer, marking a note closed without a receipt as exactly that — a strengthening. Two things I would want fixed as review notes rather than parking: the PR body says a note names the commit it was raised on, yet notes 217/220/224/226/229 carry `commit: null` while the page text promises that provenance; and the harvested `answer.by: ejhong` renders as "answered by ejhong", although the excerpt itself says the operator wrote it on the founder's direction — the display should name the operator as author so an AI reply is not read as a human one (§3.15). Neither rises to fabrication or misattributed stamping on the evidence shown.

The review-note harvester can mislabel a comment as the constitutional answer merely because its author is on `ANSWERERS`; `answerFrom` does not verify that the comment answers the objection. The added test positively demonstrates the accompanying provenance description is false: despite a later passer-by comment, it returns the earlier founder comment, while the page and YAML header call the receipt “the last comment on the issue.” This conflates the source fact—an authorized account posted a comment—with the editorial inference that the objection was answered, weakening the on-record answer requirement in §3.15 and producing an inaccurate provenance label under §§3.8–3.9. Require an explicit structured answer marker and describe the receipt as the last qualifying answer, not the issue’s last comment. The Spend copy also overstates coverage by saying every paid call is ledgered while the added historical records describe pre-2026-09-09 Arbiter calls as unmetered and the UI itself calls them unpriced; the claim should be date-bounded and disclose incomplete historical coverage under §3.13.

The change is a clean UI and domain refactor that transforms `/panel` into `/operations`, surfacing operational status, budget and spend ceilings, schedule, sittings, and the queue of review notes without mutating underlying epistemic ledgers. AGENTS.md is not modified, no citations or claims are added under `content/`, and mechanical citation checks report no new identifiers. Read-side helpers for budget, spend, configuration, and scheduling are cleanly relocated to the domain layer while preserving all validation and fail-closed checks. Harvested arbiter records and review note YAML files accurately record provenance, including issue receipts and commit references, without presenting unratified material as ratified or weakening §3.15 checks.

This is a UI/domain refactor plus governance harvest: /panel becomes /operations, domain readers for config/spend/budget/schedule move out of the pipeline so the page derives the same next-action and caps the loop uses, and harvested arbiter verdicts plus review-notes land under governance/. No AGENTS.md edit, no content/ citations (verification empty), no run records, and the OMITTED FILES list is empty. Review-note harvest strengthens §3.15 inspectability: last comment by founder/bot is the answer receipt, closed-without-answer is shown as such, and notes name the commit they were raised on. Checks are not loosened; disagreement, spend, kill-switch state, and the queue stay public.

This is a UI/infrastructure change with no AGENTS.md edit, no content/ additions, and no citation identifiers (the verification section is empty), so the fabrication, constitution, and confidence rules are not engaged. The large block of governance/arbiter/*.yaml files are append-only, explicitly labeled verbatim harvests of public PR comments with promptVersion, judgedAgainst commit, and per-seat reasoning preserved — disagreement is displayed, not hidden, and the change strengthens §3.15's inspectability rather than weakening any check: the harvest now restricts 'answers' to the founder's login or maintenance bot, paginates so no note falls off, and records a closed-without-answer note as exactly that, directly implementing review notes #229/#230. The schedule/budget/config/spend changes are verbatim moves from pipeline to domain (the new schedule.ts and the deleted pipeline/next.ts bodies match line for line), with fail-closed parsing (a bad review-note file throws at build; unparsable run records are skipped, never repaired), consistent with §8's no-silent-repair rule. The one item I cannot independently verify — that the harvested issue states and comment texts match what GitHub actually holds — is external data beyond this packet; the records carry URLs and dates, are internally consistent, and postdate my training, so per temporal humility their unfamiliarity is not evidence of fabrication.