No AGENTS.md edit; the only check-touching changes strengthen or extend gates rather than weaken them (promote-core now requires a quoted source passage in evidence sourceStatements; the intake DecisionSchema narrows the agenda-only 'proposed/scored requires proposal' rule but adds an equivalent, stricter rule for the new research-proposal stage plus stamp-consistency checks). The new research envelope validates whole prospective ledgers through the production loader and explicitly forbids AI from assigning human_reviewed, 'verified' library status, or overwriting frozen study criteria (§3.15, §3.14). The Deep Memory proposal is recorded as stage 'research-proposal'/'proposed' in append-only intake, with ai_verified/ai_extracted labels, an honest locator that self-describes as a retrieved-text character range and not a printed-page position (§3.8), direction 'context' with strength and limitations, an independence group tied to the physical pillar rather than the URL (§3.5, §3.10), and the Met's HTTP 429 recorded as an explicit miss rather than silently dropped (§3.11). Source statement vs editor inference are separated (§3.9), and the blog-not-primary-report limitation is stated (§3.7). Images are labeled: the cover as AI-generated with its prompt, the two plates with photographer, license, Commons source URL and captions distinguishing depiction from interpretation. The tooling reported no citation identifiers under content/ because the new URLs live in proposals/; the dainst.blog Pillar 43 post and Met object 322614 are consistent with their described content and I see no internal contradiction in the provenance.

The Pillar 43 bundle remains explicitly proposed and AI-extracted, labels its source `ai_verified` rather than human-verified, and records the URL, retrieval time, hashes, extractor, exact character locator, quotation, model, runId, and date. It separates the source statement from editorial inference, classifies the evidence as context, and ties independence to the underlying Pillar 43 object rather than treating another publication as independent evidence. The Met retrieval failure is retained as an HTTP 429 miss and does not become an automatically verified evidence record. The proposal mechanism validates linked records against the production loader, binds them to case and input hashes, preserves prior-record hashes and immutable intake decisions, and cannot directly modify canon or claim ratification. No AGENTS.md amendment, fabricated human review, silent assessment replacement, confidence-material publication, or weakening of the publication gate is present.

The change introduces an intake and validation pipeline for research change proposals without altering AGENTS.md or weakening existing publication gates. Proposed ledger edits are marked strictly as ai_extracted and ai_verified, explicitly forbidding automated assignment of human review or verified library status. Source locators distinguish character offsets in retrieved web text from printed page numbers, separating observation from inference and checking independence across shared objects. The added deep-memory topic files and intake records properly document AI-generated artwork, real public commons photographs, and retrieval receipts without presenting unratified material as published consensus.

The PR adds a common research-proposal envelope and a bounded manual source-reader that validates Source/Claim/Evidence bundles through the production loader, with case/input hashes, append-only intake, and no silent canon writes. The Deep Memory trial proposal stays staged as proposed/partial (not published), labels verification ai_verified, uses character-offset locators on retrieved text, separates sourceStatement from editorInference, marks independence on the shared Pillar 43 object, and keeps direction as context only. Failed runs, Met 429, and budget receipts are retained honestly; images carry AI-generated and CC licenses/credits; promoter now requires a quoted passage. No AGENTS.md edit, no fabricated citations (none added under content/), and checks are strengthened rather than weakened.

The diff edits no protected file (AGENTS.md is untouched), and the check machinery is strengthened rather than weakened: the promoter now requires a verbatim quoted passage, AI proposals are blocked from minting human_reviewed/verified states, proposals bind to case and founding-input hashes, and nothing publishes without the ordinary gate — the trial records are explicitly labeled proposed/ai_extracted/ai_verified, with failed runs (HTTP 400, 429, discarded partial run) retained honestly rather than converted into findings. Provenance is exact and honestly labeled: the Pillar 43 quote carries retrieved-text character locators explicitly marked 'not a printed-page locator' with response/text hashes, which matches the well-known Dietrich Tepe Telegrams post (2016) describing the headless, ithyphallic figure; the Met relief 32.143.7 record and CC-licensed Wikimedia images are plausibly attributed and nothing contradicts them, so temporal humility bars a fabrication vote. The source-reading adapter is a manual, bounded script rather than live AI calls or autonomous research in the site itself, with explicit spend/token/time limits and no hidden retries, and the same-vendor reading check is candidly labeled as not constitutional concurrence — nothing is presented as a ratified judgment. No content/ records changed, so nothing published is silently altered; the incubating topic remains incubating pending a separate gated PR.