The committed content is a question-only opening: claims/evidence/sources/research/dispositions/inputs manifest are all empty arrays, the edition sets `assessment: null`, `previous: null`, `featuredClaimIds: []`, and its `model` field honestly says no model was called and that the operator (claude-fable-5-1) wrote it — no assessment is presented as ratified, and no evidence, locator or priority is invented. The only external citation added under content/ is https://ejhong.github.io/birdmen/, which the tooling resolves (HTTP 200), and it is labeled as the founder's research site, not as evidence. No copy of the founder's pages is committed: the DECISIONS entry records that the GPT seat's confidence objection parked the earlier `published:`-footing copies and that they left the change, with the founder-drop door left as the compliant route — so §3.15's permission rule is honored rather than dodged. AGENTS.md is untouched; the test changes only exempt the assessment-bearing invariants for a case whose edition adopts nothing, while adding a new test asserting an opening derives no standing — this narrows nothing for assessing editions. The spend rows for the two discarded rehearsal runs are retained ("the bill is the bill"), which is honest bookkeeping. One review-note-level discrepancy: the PR body asserts the pages "came in through the inbox" and were registered as DEEP-IN001–IN004 with extractions under research/deep-memory/, none of which appears in the diff; the DECISIONS and history entries correct this (inputs "to come", none committed), so I read the body as stale prose rather than a false record, but I would want the body reconciled with the committed state.

The opening article says the founder’s pages are founding inputs “taken in through the inbox,” while `inputs/manifest.yaml` is empty, the edition records an inputs hash for no inputs, and `docs/DECISIONS.md` says “Once the drop lands.” The PR body likewise says the pages came in and were registered by run `…191337`, whereas the complete diff contains no registered inputs or processed drop and explains that those outputs left the change. This publicly misstates when and how the inputs entered, contrary to exact, reconstructable provenance and accurate change history under §§3.8, 3.14, and 3.15. The text and PR record should instead describe the drop and intake as future, while identifying the current summary as derived from the resolved public URL, or include the actual founder-drop records, manifest entries, and matching input hash before claiming intake occurred.

The change initializes the 'deep-memory' case question-only with an empty ledger and a null assessment opening edition, inventing no claims, evidence, or reviews. The external research URL resolves mechanically with HTTP 200, and no founder-supplied text is committed without provenance or proper drop procedures, respecting §3.15. The supporting UI and pipeline adjustments cleanly handle question-only editions without weakening existing validation or ratification checks.

The diff opens DEEP-001 as a true question-only case: empty claims/evidence/sources/research/dispositions/inputs, assessment null, featuredClaimIds empty, previous null, and an operator-written opening that grades nothing. No AGENTS.md edit; no citations or locators are invented (the sole added URL resolves); founder page copies are not committed after the prior confidence objection, with intake left for the founder-drop door. History/edition stamps honestly attribute the operator and the empty-ledger basis hashes. UI/test/inbox changes only teach the opening shape and fix the processedName locator; they do not weaken §3.15 checks.

The committed diff contains no founder-supplied copies: the inputs manifest, claims, evidence, sources, research, and dispositions are all empty, and the DECISIONS.md entry in the diff records that the operator's byte-identical copies were removed after the GPT seat's confidence objection, with the pages now entering only by the founder's own drop under the standing direction — which is exactly what §3.15's permission rule requires, so the confidence hazard the body describes was resolved rather than merged. The URL ejhong.github.io/birdmen resolves, and the only committed citation is an external-research link to that public site, honestly labeled. The opening edition is stamped 'model: none — written by the operator', adopts assessment: null, features nothing, and its article says plainly that nothing is graded and nothing on the page is evidence — honest labeling and calibrated uncertainty. The test skips are narrowly scoped (isOpening) to checks that are inapplicable when no assessment exists, with a new test covering the opening's invariants, so this is extension, not check-weakening; the processedName fix actually tightens locator accuracy under §3.8. One note for the record: the PR body still describes the intake as having 'registered them as DEEP-IN001–IN004' and the pages as already come in, which the diff contradicts (research.yaml is empty, DECISIONS says the copies and run outputs left the change) — the committed history and DECISIONS entries are accurate, but the body text should be corrected to match what actually merges. The rehearsal spend rows are retained rather than hidden, which is honest accounting.