The diff adds only two editions, the draft assessments they adopt, and their run, comparison, errors and spend records; no claim, evidence or source record, no check, and not AGENTS.md is touched, and the tooling found no new citation identifiers under content/. Both assessments carry runId, producedBy, model, date and promptVersion, and are marked role: draft and humanReviewed: false; the PR states that standing shows as not yet ratified pending blind checks, so no draft is presented as ratified (§3.15, §4). The only judgment change is disclosed in the rationale, the verb's measured line and the PR body, each with its reason and basis records (§3.14): GEO-C913 and GEO-C918 move from well to provisionally supported (GEO-E057 and GEO-E058 on parent GEO-C004), and GEO-C502 moves to unresolved (GEO-E002 flags it unverified); the earlier run that moved a verdict silently was discarded, and its spend is still recorded. The provenance I could check is internally consistent: each run.yaml's usd, input tokens (including 384 cache-read) and output tokens equal its spend rows; the discarded runs total $1.83 and $1.76; seat presentation orders match the PR's first/second counts; errors.md matches the repairs, including the three ']' closers visible in the geopolymer reply.json. Every article claim link is featured and graded, and the figures track the assessments (216 cm³/h, PI 8.3 vs 3.0, 11/11 vs 4/11 sites). Unverified sources and editorial inferences are labeled as such (the second essay, Klemm & Klemm, 'This edition reads that as a lead'), per §3.9. The one omitted file is the superseded first vasocomputation reply, a working file whose shape (23 claim assessments, ~3,029-word article) matches the documented failure and whose adopted successor is fully visible, so it cannot change the outcome. Remaining nits are ordinary imperfection for the next drafter, not violation: for example, GEO-C594 is stated flatly though graded provisional and unaudited beyond Giza.

The visible installed assessments are honestly marked as AI-generated drafts, identify model/run/date/prompt version, and state that standing still requires blind checks; the comparison records also explicitly say they judge the tellings rather than the verdicts. The articles generally preserve uncertainty, competing accounts, evidence dependence, and the distinction between source reports and editorial inference. However, proposals/2026-09-30-edition-vasocomputation-071302/reply.json is omitted and is a newly committed public-git artifact containing a full article and assessment, so I cannot verify that it contains no fabricated locator, mismatched provenance, or confidential supplied material. I would need that complete omitted file, or equivalent repository-tool verification of its substantive claims and provenance, before finding the whole change compliant.

The pull request adds two shortened editions and their corresponding assessments for 'geopolymer' and 'vasocomputation' without modifying the underlying claim/evidence ledger or editing AGENTS.md. Both editions were produced via documented pipelines with repair rounds recorded, compared against their incumbents by five-model panels (where all seats preferred the candidates), and explicitly labeled as unreviewed AI drafts (`humanReviewed: false`, `role: draft`). Exact provenance stamps, run records, spend logs, and basis hashes are preserved intact, fully complying with constitutional standards.

The change adds two draft editions and their assessments, plus spend rows and proposal artifacts for the completed runs (and two discarded ones). Both assessments are stamped role: draft, humanReviewed: false, with model/runId/date/promptVersion and ledgerHash basis; no claim, evidence, or source records are touched. Case verdicts stay unresolved; the three GEO claim moves are disclosed in the rationale as applying the source-statement-alone rule, with provisional/unresolved grades and disclosed parent-evidence bases. Comparison seats preferred both candidates 5–0; repair rounds are recorded. No AGENTS.md edit, no fabricated citations (none added), and no confidence material.

The change adds only draft assessments, editions, spend records and working files; no AGENTS.md edit and no new citation identifiers appear (verification found none), so no fabrication check is triggered by the diff. The assessments and editions are stamped with runId, model, date and promptVersion, carry humanReviewed: false and 'NOT human reviewed' headers, and are labeled AI drafts — no unratified material is presented as ratified; the PR body explicitly states both cases stand unratified pending blind checks, and the comparison files self-describe as AI judgments. Verdict movements (GEO-C502, GEO-C913, GEO-C918) are honest disclosures that claims rest on source statements or parent-claim evidence, moving toward lower confidence rather than overstating support, consistent with §3.8's unverified-locator rule and the rule on claims held on source statements alone. Provenance is internally consistent: each run.yaml 'wrote' list matches the files added, and the spend ledgers sum to the run.yaml costs (4.331452 and 2.899427), with the two discarded runs' spend honestly included; the raw malformed reply.json is preserved unmodified as a working record rather than silently cleaned. The repair rounds (errors.md) show the budget and marker checks being enforced, not weakened, and I found no confidence material, fabricated locator, or check-dodging reclassification in what is shown.