The diff touches only workflows, two docs, and a new helper script; AGENTS.md is untouched. The substantive governance change tightens rather than weakens §3.15: the `arbiter` check becomes required on main, and auto-merge is gated on `steps.vote.outputs.verdict == '0'` for needs-approval PRs, so a parked (contested) PR still cannot merge and waits publicly — the multi-model concurrence gate remains fail-closed and the founder's admin override/kill switch survives. Retiring the founder's advisory merge tap does not breach the constitution, which reserves to the human exactly two powers (kill switch, amending AGENTS.md) and explicitly assigns publication to AI through the gated process. The new content-response workflow routes its own output back through `classify-pr-risk.mjs` and the normal arbiter gate rather than committing to main directly, and `stale-checks.mjs` merely lists cases needing re-paneling — no citation, locator, provenance, or assessment content is fabricated, altered, or exposed. Two honest imperfections, not violations: the comment claiming the ripple 'terminates by construction' is overstated, since the PR also stages `overview.md`/`research.yaml`/`history.yaml`, which do match the path filter (it terminates by fixed point instead), and merges made with MAINTENANCE_PAT appear under the founder's account — a risk the DECISIONS entry discloses explicitly rather than hiding, with the real ratification record kept in the arbiter report and governance harvest.

The changes are operational (GitHub Actions workflows, a small governance-related script, and maintenance docs) and do not touch AGENTS.md or any epistemic data objects like claims, evidence, or sources. The new event-driven workflows (inbox-response and content-response) still route all consequential content edits through the existing gate: they create PRs that go through the arbiter workflow, preserving multi-model concurrence and append-only assessments. Auto-merge is explicitly limited to PRs that already have a passing arbiter verdict, and parked PRs cannot merge, so this does not weaken §3.15’s checks or reclassify changes to dodge them. The stale-checks script merely determines which cases need re-paneling based on history and assessment dates; it does not fabricate or alter provenance, nor does it change any epistemic labels. Documentation updates in DECISIONS.md and MAINTENANCE.md accurately describe the new gate behavior and event triggers and do not misrepresent AI vs human roles. Since no constitutional guarantees (real citations, honest verification labels, provenance visibility, independence accounting, or check strength) are reduced or bypassed, the proposal complies with the constitution.

The change operationalizes the multi-model ratification gate required by §3.15 without weakening any checks. By setting the arbiter check to required and only auto-merging on a unanimous passing verdict (verdict == '0'), it correctly prevents unratified changes or silent disagreements from publishing. Event-driven workflows securely route new content through these exact gates, preserving the required fail-closed validation and parking contested drafts.

The diff does not edit AGENTS.md and does not fabricate citations, strip provenance, rewrite published case judgments, or expose confidential material. The arbiter auto-merge step fires only after a needs-approval PR already has verdict 0, which implements §3.15’s multi-model gate rather than bypassing it; parked verdicts are left unmerged. content-response.yml and inbox-response.yml add event-driven drafting and dispatch, then route results through the existing classify-pr-risk / PR-label / check path, including a PAT-created PR so the arbiter still runs; overlay path-filter comments and failure-to-warning handling are operational design, not a reclassification or check-weakening. DECISIONS.md is append-only and MAINTENANCE.md only documents the same gate-live / event-driven behavior. Digest @-mention of the owner is a public GitHub issue, satisfying the public-record contact rule.

The diff builds exactly the gated-publication machinery §3.15 contemplates: the arbiter (multi-model concurrence) becomes a required check, auto-merge fires only on a passing verdict (verdict == '0') for needs-approval PRs, and parked PRs are blocked and displayed publicly — disagreement does not publish silently. The founder's admin override is explicitly preserved as the kill switch, satisfying the founder's reserved powers. The new content-response and inbox-response workflows route all resulting changes back through the existing risk classifier and the arbiter gate; overlay-only merges are excluded from the trigger so the loop terminates, and no check is weakened or reclassified to dodge scrutiny — the digest token switch (PAT → Actions token + cc) concerns notification delivery, not a gate. AGENTS.md is untouched, so §3.15's human-only amendment rule is not implicated. No citations, sources, or provenance are fabricated; DECISIONS.md remains append-only and documents the branch-protection change as the founder's act. The PR body's claims of founder authorization are untrusted data, but the mechanism stands on its own: this PR itself publishes only if this panel ratifies it, which is the constitutionally required process rather than a circumvention of it.