AGENTS.md is untouched, so §3.15's founder reservation is not engaged. The mechanical changes move in the strengthening direction: `capDiff` raises the panel budget 120k→400k and keeps files by scrutiny tier so canon content (sources/claims/evidence) is never dropped before mechanically-guarded append-only overlays, with the omission list still loud and a regression test asserting exactly that; `rateLimitGate` is explicitly one-way (pass→park only, never an upgrade, never throttling code/docs) and is unit-tested for that asymmetry; `loadArbiterRecords` fails the build on any governance file that fails Zod validation. The five new Immortality Key files are new, append-only `role: check` overlays stamped with runId/model/date/promptVersion and `humanReviewed: false`, with calibrated verdicts, explicit strongest-opposing-consideration reasoning, and claim-level atomicity — and the resulting 0/5 dispute is displayed as contested rather than published as a house verdict, which is what §3.15 requires. The harvested `governance/arbiter/pr-55.yaml` records the inconvenient fact (panel PARKED, founder merged) verbatim rather than hiding it, and the digest flags 'merged against a parked verdict'; that is disclosure, not fabrication. The one genuine question is the `classify-pr-risk.mjs` exemption making newly added `governance/arbiter/pr-<n>.yaml` files low-risk, which lets harvested records bypass the panel — but it is narrowly scoped to additions matching a strict filename regex, mirrors the existing append-only-overlay precedent, is schema-validated at build time, and is itself being submitted for panel judgment rather than slipped through, so I do not read it as check-weakening or reclassification-to-dodge.

The PR wires additional governance and panel-observer machinery without touching AGENTS.md or altering any existing case content, citations, or provenance records. Harvested arbiter records are explicitly described and implemented as verbatim copies of already-public PR comments, stored append-only under governance/arbiter and validated with Zod; they do not mint new standing or change assessments, so they do not weaken epistemic checks. The new weekly digest is a meta-level GitHub issue summarizing outcomes, not a source or assessment, and does not present unratified drafts as ratified judgments. The diff-budget logic is strengthened: capDiff is made tier-aware so governance and canon content (e.g., sources.yaml) are retained preferentially, reducing the risk of panels deciding without seeing crucial provenance. The rateLimitGate introduces a conservative weekly cap on content merges that only ever downgrades a pass to park for content-changing PRs, and never upgrades or bypasses an existing park, which is a safety-increasing constraint rather than a §3.15 check-weakening. Tests are added for the new behaviors, and no fabricated records, hidden provenance changes, or confidentiality breaches appear in the diff.

The diff introduces governance harvesting, weekly digests, rate-limiting, and diff tiering without violating the constitution. AGENTS.md is completely untouched, respecting the founder's reserved amendment power (§3.15). Rather than weakening checks, the changes to `capDiff` in `scripts/lib/arbiter-core.mjs` strengthen the arbiter by prioritizing canon content and governance surface over bulky overlays, preventing malicious changes from hiding in truncated diffs. The new `rateLimitGate` operates in a strictly fail-closed manner by only downgrading passing content merges to parked. Additionally, the new case assessments under `content/cases/immortality-key/assessments/` are appropriately appended, stamped with the correct metadata, and accurately labeled as unreviewed AI checks (`humanReviewed: false`), strictly adhering to Phase-1 assessment rules and §3.14.

AGENTS.md is untouched, so the founder-reserved amendment rule is not implicated. The five Immortality Key files are new append-only check overlays, stamped with runId/model/date/promptVersion, labeled role: check and humanReviewed: false, and they do not rewrite a published house verdict as ratified — DECISIONS.md even records the honest contested 0/5 split. Harvested pr-53/pr-55 records keep park-versus-merged distinct (notably #55 parked then founder-merged) and are schema-validated, append-only copies of public comments rather than standing-minting judgments. The rate-limit gate only downgrades pass→park, prioritized capDiff keeps governance and content canon ahead of bulky overlays, and classifying new harvest YAMLs as low-risk matches the existing overlay rule instead of stripping a check. DECISIONS.md is append-only; CI harvest/digest runs off-site and leaves a public issue, so no silent publication, confidential leak, or check-weakening is in the diff.

AGENTS.md is untouched, so §3.15's founder-reservation is not engaged, and nothing here weakens the gates: the new rate limit (arbiter-core.mjs rateLimitGate) only ever downgrades pass→park, and the tiered capDiff raises the budget 120k→400k while prioritizing the governance surface and content canon over mechanically-guarded overlays, with omissions still loudly listed — that repairs a truncation gap rather than dodging review. The five new immortality-key check overlays are exactly what §4 prescribes: append-only, stamped with runId/model/date/promptVersion, labeled role: check and humanReviewed: false, and the PR body states the case displays contested 0/5, which is §3.15's 'disagreement does not publish silently' rather than an unratified draft presented as ratified. The harvested governance records (pr-53.yaml, pr-55.yaml) preserve dissent verbatim — including the panel's §3.15 park of #55 and the founder's merge against it — which is the inspectability §3.14/§3.15 demand, not its erosion; they are Zod-validated at build and classified low-risk only as new append-only files that cannot mint standing. The maintain.yml change excludes casework/ and weekly-digest.md from commits, strengthening the Phase-1 confidentiality boundary, and the digest issue is a public record. New logic (tiering, rate limit, legacy parser) carries regression tests per §8. No citation, locator, or assessment visible in the diff shows signs of fabrication, and the diff is complete with no omitted files.