Nothing in the diff touches AGENTS.md, content/, or any citation — the mechanical verifier found no identifiers, and the changes are confined to the workflow schedule, governance state file, two docs, and one constant in scripts/arbiter.mjs. Turning on the weekly cron and setting `governance/operation.yaml` to `state: live` are governance toggles the constitution reserves to the founder's kill switch, and the file is stamped with state, date, `by: founder`, and a reason naming the direction's channel (session, 2026-09-09), which satisfies the provenance discipline for founder directions in §3.15. The DECISIONS entry is dated, channel-stamped, and honest about limits ("What has not yet run in CI: a check, a report, a draft, a verify"). The AUTOMATION.md edit deliberately preserves the pre-run bullet as the record rather than overwriting it, which is the right append-style handling. The one item worth a note is the GATE_EPOCH move in arbiter.mjs: it clears the content-merge throttle's window and is at least the fourth such bump. On the face of it, though, this is a rate throttle, not one of the checks enumerated in §3.15 (schemas, dangling IDs, provenance completeness, real citations, multi-model ratification, parking on contest), and the comment gives a coherent non-dodging rationale — the spent window consisted of hand-run, hand-merged landings that the autonomous lane should not inherit — with the direction dated and attributed. The narrow-veto and single-seat-park language reproduced in operation.yaml matches §3.15 as ratified, so no check is described as looser than it is.

The `scripts/arbiter.mjs` change advances `GATE_EPOCH` specifically to exclude fourteen recent content landings after the 10-change throttle had parked the chain. That clears a fail-closed gate precisely because its limit was reached, rather than waiting for the window to expire, and therefore weakens or dodges the check prohibited by §3.15. Recording a founder direction in the comments does not itself amend that constitutional prohibition. Keep the existing epoch and wait for the window to clear, or implement a prospective, auditable distinction between manual and autonomous landings without erasing already counted events.

The diff implements the founder's direction given in session on 2026-09-09 to activate the autonomous maintenance chain on a weekly cron schedule, updating governance/operation.yaml to 'state: live', resetting the arbiter throttle's epoch past the day's manual landings, and updating operational documentation accordingly. AGENTS.md is not modified, no confidence materials or citations are involved, and all changes faithfully record the founder's operating directions in DECISIONS.md and config per §3.15.

The diff turns the scheduled chain on, sets operation.yaml to live, advances the content-merge throttle epoch past the construction-day hand merges, and records the founder’s 2026-09-09 in-session direction in DECISIONS/MAINTENANCE/AUTOMATION. AGENTS.md is untouched (§3.15). No citations, sources, locators, or confidence material are added. The GATE_EPOCH bump only clears the weekly content-merge count; it does not loosen fabrication, confidence, constitution-edit, or multi-model ratification checks. Spend bounds and kill-switch path remain as documented.

This is an operations/infrastructure change — enabling the weekly cron, flipping governance/operation.yaml to live, and moving the Arbiter's throttle epoch — with no content/ additions, no citations, and no edit to AGENTS.md, so the fabrication, confidence, and constitution rules are not triggered. The one rule that needs scrutiny is §3.15's prohibition on weakening checks: the GATE_EPOCH bump in scripts/arbiter.mjs does clear a throttle window that had parked #214, but the throttle is a merge-rate limiter, not one of §3.15's epistemic checks (fail-closed CI, panel ratification, provenance), all of which remain intact and are described as still running on every PR. The epoch move is explicitly attributed to a founder direction in session on 2026-09-09, recorded in the append-only DECISIONS.md with date and channel, and mirrors the prior founder-directed epoch bump of 2026-09-03 — it is a documented recalibration to clear a known miscount from fourteen hand-merged landings that carried no trailer, not a reclassification to dodge ratification (#214 was itself panel-judged 5/5 before the founder merged it). The kill-switch documentation is preserved and strengthened (paused state or cron removal both stop it, displayed on the pages), and spend is bounded by config/budget.yaml. The unverifiable claims in the PR body (that the founder directed this in session, that #214 ran and merged) are provenance assertions about an operating record I cannot independently inspect, but nothing in the visible material contradicts them, the account is internally consistent, and no omitted file could hide a content-level violation since the diff is complete and touches no content.