The diff adds an immutable, append-only edition chain (content/cases/deep-memory/editions/edition-2026-09-06-opening.yaml) whose article is byte-identical to the deleted overview.md, with the essay's two citations and the {claim=MEM-C001} reference preserved; loadCase now rejects having both overview.md and an edition, so there is one article authority and no silent loss of provenance. Assessment handling strengthens rather than weakens honesty: assessment is explicitly null ("This edition leaves the question unassessed"), editionDraft throws if a pinned assessment's hash changed, EditionProposalSchema refuses runs with humanReviewed or role "check", and the original model/date are referenced rather than copied into an editor's stamp (§3.14, §3.15 respected). Gates are not relaxed: MAINTENANCE.md explicitly adds "new or changed editions" to the needs-approval lane, the panel stays enabled, thresholds unchanged, and ledger changes invalidate old receipts plus surface an editionStale notice. The reassess-changed and reconcile-contested early-returns route work into edition proposals with the incumbent retained and dissent still derived, which is a workflow reroute, not a check bypass. AGENTS.md is untouched; the DECISIONS/AUTOMATION/DATA_MODEL updates describe limits candidly. The only failing citation check is the Met URL with HTTP 429 (rate limiting, not nonexistence), and it is pre-existing text moved verbatim, not a newly asserted source.

The change preserves the existing Deep Memory essay byte-for-byte in a versioned edition, explicitly records that no assessment was made, and keeps the edition’s model, runId, timestamp, prompt version, hashes, and predecessor chain visible. Assessment references are hash-bound to the original assessment record, while later unadopted drafts cannot silently alter the displayed judgment; ledger changes invalidate standing and produce a stale-edition notice. The DAI URL resolves, and the Met URL is an identical relocation from the deleted overview; its HTTP 429 is an access-rate failure rather than positive evidence of a fabricated locator. The diff does not edit AGENTS.md, expose confidential material, or weaken the publication panel, and it adds validation for stale inputs, dangling references, assessment identity, and edition history.

The change introduces an immutable, versioned edition mechanism that pairs an essay, claim selection, and assessment reference without mutating past records or fabricating data. Deep Memory's original overview text is preserved byte for byte in its initial edition, keeping the question explicitly unassessed and retaining its existing photographic plates and citations. All edition references, input hashes, and load-bearing claim requirements are strictly validated in domain loading and schema checking. Maintenance documentation and workflow policies ensure editions remain classified under the consequential-content gate requiring multi-model arbiter concurrence.

The diff introduces immutable edition manifests that bind essay text, claim selection, and an exact assessment reference (or explicit null), preserving original assessment authorship rather than copying or relabeling it, which aligns with §3.14 and §3.15. Deep Memory’s opening is moved byte-for-byte into the first edition with assessment: null, plates and MEM-C001 intact, honest model stamp “no model call,” and no fabricated sources, locators, or judgments. Edition pages and UI copy disclose that archive links/captions use the current ledger and that unassessed openings stay unassessed; catalog/explorer wording no longer overclaims absence of evidence. Authoring validates prospective cases, refuses silent overwrite of assessments, keeps AGENTS.md untouched, and leaves the multi-model gate and thresholds unchanged; CITATION VERIFICATION shows the Tepe URL resolving and the Met URL as HTTP 429 (not non-existence), matching the preserved incumbent text.

The edition manifest preserves the Deep Memory opening byte-for-byte (verified against the deleted overview.md), keeps both plates and the MEM-C001 claim reference, and honestly records assessment: null with model 'Aletheia incumbent assembler (no model call)' — no fabricated judgment or human review impersonation (§3.15, §4). The dainst.blog URL resolves; the Met URL's HTTP 429 is a rate-limit on a URL preserved verbatim from already-published content, not positive evidence of fabrication. The edition chain is validated (predecessor hashes, no forks, load-bearing claims must stay selected), assessment authorship is pinned by runId+hash rather than relabeled, and editions are explicitly added to the needs-approval arbiter lane, which strengthens rather than weakens the gate (§3.15). Skipping legacy reassessment/reconciliation for migrated cases does not silence disagreement: standing is still derived from blind checks, stale editions show a visible notice, review receipts invalidate on ledger change, and all runs remain in append-only history (§3.14). The claims-explorer prose change removes a page-level blanket sentence but per-record ai_extracted provenance labeling is untouched, so §4's display requirements still hold. No AGENTS.md edit is present, and no source, evidence, claim, or assessment record is modified.