The only new content citation (the DAI 'Tepe Telegrams' Pillar 43 post) RESOLVES, and MEM-C002/MEM-E002 use it honestly: reviewState ai_extracted, direction 'context', an exact retrieved-text locator explicitly labeled 'not a printed-page locator' (§3.8), limitations noting it is a blog summary rather than a primary excavation report, and the same-object independence group reused so the two Pillar 43 observations are not double-counted (§3.10). The new edition and its assessment are stamped model/runId/date/promptVersion with humanReviewed:false, role:draft, and the essay itself states it is an AI-generated assessment, not a reviewed human conclusion (§3.15, §4); the winner is only 'prepared' for the ordinary PR arbiter plus five fresh blind checks, all of which are recorded here including the seat that ranked the incumbent first. The earlier failed and contested comparison receipts, rejected replies, seat errors and dissenting reasons are preserved append-only in proposals/intake with their original version-1 tallies rather than being reinterpreted (§3.14). The workflow edits remove a cost-saving supersede/close path in favour of resting while a pending batch (including a park) stays open — that preserves rather than weakens the gate, and no arbiter threshold or vendor-seat requirement is lowered. AGENTS.md is untouched, and the dropped IMG-MEM-P02 plate is explained in rationale and history rather than silently removed.

The added Pillar 43 claim is a bounded observation tied to a resolving URL, an explicitly described retrieved-text locator, retrieval hashes, and a separate source-reading check; its Evidence record distinguishes the source statement from editorial inference and assigns an explicit context direction. MEM-C001 and MEM-C002 share the same independence group, and the selected edition repeatedly warns that one object and one source do not constitute independent support or establish cultural transmission. The assessment remains calibrated as unresolved, distinguishes local credibility from low diagnosticity, presents serious alternatives, and labels itself AI-generated and not human-reviewed. Failed drafts, the original contested comparison, all ballots and dissent are preserved append-only, while candidate preparation rechecks recorded receipts, current inputs and the ordinary publication gate; no AGENTS.md amendment, fabricated citation, silent assessment replacement, or check weakening is shown.

The diff implements automated edition drafting and independent multi-vendor candidate comparison without weakening publication checks or altering AGENTS.md. In content/cases/deep-memory/, the added claim MEM-C002 and evidence MEM-E002 record narrow physical observations from an existing verified source, honestly labeled with exact retrieved-text character locators (§3.8) and maintaining the shared object's independence group (§3.10). Source statements and editor inferences are cleanly separated (§3.9), and the evidence direction is explicitly marked as context (§3.5). The new edition and draft assessment explicitly disclose their AI-generated, unreviewed status and hold the case verdict as unresolved (§3.13, §3.15). Complete receipts—including failed attempts, rejected replies, and seat-by-seat comparison rankings—are preserved in proposals/intake, and case history entries are strictly append-only (§3.14).

The diff does not edit AGENTS.md and adds Deep Memory content with honest AI labels (ai_extracted, humanReviewed: false, role draft/check), append-only assessments, and a single RESOLVES URL used only for bounded Pillar 43 observations. MEM-C002/MEM-E002 keep exact retrieved-text locators labeled non-page, direction context, shared independence group, and source-statement vs inference separation; the adopted edition leaves inheritance unresolved and omits the unledgered Met/Assyrian comparison. Edition-cycle receipts preserve failed drafts, the contested v1 ballot, and v2 ranking-vs-status judgments; preparation still requires the ordinary gate and fresh blind checks. Workflow backpressure rests while a respond PR is open rather than silently dropping parks, and no concurrence threshold, citation rule, or check is weakened.

The diff does not touch AGENTS.md and keeps every publication gate intact: the edition comparison is explicitly a drafting decision, with the ordinary PR arbiter, fresh blind checks, and fail-closed validation still gating publication ('preparation is not publication', tests confirm a forged tally and a violates-vote park the outcome). The only added content citation (the DAINST Tepe Telegrams URL) mechanically RESOLVES, and all new records carry honest provenance: AI-extracted reviewState, retrieved-text character locators explicitly labeled 'not a printed-page locator', separated sourceStatement/editorInference, a single shared independence group for the one pillar, and 'humanReviewed: false' AI-draft labels. The removal of the incumbent's unledgered Met link and the unbacked Assyrian plate is an epistemic improvement (§3.7/§3.8), with the rationale and prior opening preserved in append-only history and intake receipts rather than silently deleted. The workflow change replaces PR-superseding logic with conservative backpressure that keeps parked PRs visible (§3.15) and does not lower any ratification threshold, budget, or blind-check requirement; contested, failed, and invalid drafts are all retained in immutable intake receipts, satisfying dissent-preservation and append-only rules. Validation claims (tests, TS, lint, build, link audit) are consistent with the diff's own test coverage of rollback, rest logic, and vote integrity.