The two new edition candidates re-adopt the incumbent assessments, carry basis hashes, producedBy run ids, model/promptVersion/date stamps, and matching run.yaml and spend.yaml entries; no assessment is presented as new or as human-reviewed. The operator's pre-publication amendments are the strongest signal of compliance rather than against it: each departure from the verb's reply.json (removal of the V1-confound sentences whose evidence record was declined at intake, the calibrated conclusion sentences) is recorded in the file header, in an `amendments` block naming the reviewer note, and in an append-only history entry, so a reader can reconstruct verb text versus operator text. The word-count discrepancies are handled honestly: the verb's estimate is left standing and labelled a guess, the measured figures (4,054 → 3,849 candidate, 3,859 as amended; 5,150 → 5,209) are appended and attributed to the operator, and the earlier single-figure annotation that contradicted the run record was itself corrected on the record. The geopolymer rationale also openly reconciles five recorded accounts against four as grouped in prose. The citation tooling found no identifiers in added content, and nothing in the articles shows a self-contradicting locator or quote; the articles repeatedly mark unread primary sources, second-hand citations and unrecorded claims rather than overstating them (§3.8, §3.9, §3.13 respected). One point I would flag as a review note, not a violation: the diff retro-stamps several pre-existing by-hand history entries with `runId: session_...` and `promptVersion: operator-by-hand-v1`, including an entry whose own text still asserts that no run id or prompt version exists for it — a new history entry explaining that stamping convention would remove the apparent self-contradiction.

The change introduces internally contradictory provenance stamps. In deep-memory/history.yaml it retroactively adds `runId: session_01AiXD4y5Y91KC7iM3uBhWmu` and `promptVersion: operator-by-hand-v1` to older entries whose existing text explicitly says “there is no protocol and no run” and “No run id and no prompt version exist for it because no protocol ran.” Relabeling a private operator-session identifier as a run ID, while asserting a prompt version where the history says none existed, prevents readers from reconstructing how those changes were actually made and checked. The same synthetic stamps are added across older deep-memory and geopolymer history entries; they should be removed or represented in semantically accurate session/operator fields, with any correction recorded truthfully rather than rewriting the old provenance.

The pull request adds two edition candidates (`deep-memory` and `geopolymer`) which both re-adopt their incumbent assessments on unchanged ledgers while refining prose and explicitly recording accounts. The operator pre-publication amendments address previous panel review notes (#292, #294, #297), calibrating uncertainty in the conclusions, scoping statements to the ledger, and precisely documenting word count differences. All operator modifications and schema-completing history attributes are fully stamped with date, actor, session runId, and promptVersion in compliance with §3.8, §3.14, and §3.15.

The diff adds two stamped edition candidates that re-adopt prior assessments by explicit hash/runId with no new assessment overlay, plus append-only history entries and spend lines; operator amendments for review notes #292/#294/#297 are recorded in the edition headers, amendments blocks, and history with session/runId/promptVersion stamps. Deep-memory prose no longer re-introduces declined non-atomic confounds as ledger-held evidence and the conclusion now states only indirect support; geopolymer conclusions are calibrated to “consistent with, not established” matching the article’s own second-hand and untested qualifications. No new citation identifiers appear (verification empty), AGENTS.md is untouched, and no confidence material or check-weakening is present. The run accounts match the written files and label the re-adoptions honestly.

Both editions are honest re-adoptions of incumbent assessments with no new assessment, and the panels' dissents (review notes #292, #294, #297) are answered on the record rather than hidden, which is what §3.15 requires of contested changes. Where the verbs' rationales misstated facts (the deep-memory word-count estimate 'roughly a sixth' and the geopolymer 'at or slightly below' claim), the operator did not silently edit them but appended clearly labelled measured corrections (4,054→3,849/3,859; 5,150→5,209), keeping the record reconstructable under §3.14's append-only discipline. The rejected V1-confound material was correctly kept out of the edition prose — the amended article says only what the ledger holds (DEEP-C027) and what the retelling account predicts, exactly as review note #292 required, rather than re-admitting a declined record's propositions as prose. No citation identifiers were added, no AGENTS.md change appears, no confidence material is published, and the editions are presented as candidates for panel ratification, not as ratified judgments. The backfilled session-id and 'operator-by-hand-v1' stamps on earlier history entries are additive, match the session id the repository itself established on the record (note #247), and improve rather than obscure reconstructability; they do not misattribute who did what. The residual inaccuracies that remain inside the preserved verb rationales are exactly the kind of ordinary imperfection the append-only annotation mechanism exists to correct, not a fabrication or provenance violation.