seat failed: Error: anthropic HTTP 529: {"type":"error","error":{"type":"overloaded_error","message":"Overloaded"},"request_id":"req_011Cec9iXkneGNk51Z3jh12L"}

The change introduces a NarrativeInput schema and per-case manifests that explicitly mark founding texts as "presentation references, never evidence" and are validated fail-closed by the loader. This strengthens provenance and anti-drift controls without altering any existing evidence records, claims, or assessments, and it preserves append-only behavior and versioning through frozen snapshot files dated 2026-09-01. No citations, DOIs, or URLs are added inside content/, so there is no risk of fabricated or misused external sources; the CITATION VERIFICATION tool finds none. The new rewrite rules in docs/AUTOMATION.md (first-edition rule, plates-survive rule) operate at the narrative/editorial layer and do not weaken epistemic checks or reclassify changes to dodge multi-model ratification; instead they enforce completeness of plates/claims when rewriting. AGENTS.md is untouched, so the human-reserved constitutional power is respected. Overall, the diff tightens provenance and workflow discipline while remaining consistent with the constitution’s requirements on honesty, explicitness, and non-fabrication.

The PR implements a structured mechanism for tracking founding narratives and research inputs via Zod-validated manifests in the `content/` folder, directly adhering to Phase-1 static data rules (§4). The new `NarrativeInputSchema` enforces fail-closed validation for file existence, preventing dangling references. The provided origins and licenses in the manifests clearly document the provenance and status of each text as presentation references rather than citable evidence, fully satisfying exact provenance (§3.8). The editorial rewrite rules added to the documentation strengthen the project's standards by strictly enforcing the retention of existing visual evidence (plates) and do not weaken any constitutional checks.

The diff does not edit AGENTS.md. It adds case-local founding-narrative snapshots, per-case input manifests, a fail-closed NarrativeInputSchema/loader existence check, and two additional Atelier rewrite laws in docs/AUTOMATION.md and docs/DECISIONS.md — none of which weaken §3 checks. Manifests label inputs as presentation references, never evidence, with origin and license fields; no Source/Evidence/Assessment records are fabricated or silently rewritten, and citation tooling found no new DOI/arXiv/URL identifiers. Research PDF paths are declared as already-committed founder work, not third-party captures. Unfamiliar 2026 names and events in the frozen overviews are not positive evidence of fabrication under temporal humility.

The change adds dated, append-only narrative-input snapshots plus manifests, a NarrativeInputSchema, and fail-closed loader validation that every manifest entry's file exists — this strengthens rather than weakens the §3.15/§4 checks, and no AGENTS.md edit is present (the docs/AUTOMATION.md and docs/DECISIONS.md additions are exactly where §7's founder-amendment note says such directives are recorded). The manifests explicitly fence inputs as 'presentation references, never evidence,' consistent with §3.6/§3.7, and exclude third-party captures, keeping only material asserted as the founder's own essays and commissioned reports — so the §3.15 confidence regime for third-party-supplied material is not triggered by what is shown. The frozen narratives are verbatim snapshots of already-published overviews containing no new citations (the mechanical verifier found none), and their 2026-dated claims are internal references to existing ledger claims, not fabricated locators; per temporal humility I treat unfamiliar 2026 events as unverifiable rather than fabricated. The transients snapshot paraphrases an unpublished Yang–Villarroel manuscript 'shared with this project,' but that text pre-exists in the base overview, the PR merely snapshots it, and there is no positive evidence in the diff that it lacks recorded permission. The schema/loader code is orthogonal to epistemic content and fails closed on dangling files. No omitted files; nothing in the visible change breaks a named rule.