The change is additive: a new images.yaml, seven image files, plate-placement markers in overview.md, a tone row in docs/IMAGE_STYLE.md, and an appended history entry. AGENTS.md is untouched. The AI-generated cover is explicitly labeled as such with model, runId, styleVersion, license and the full generating prompt recorded, and its own text forbids depicting any real vessel or result — so no AI artwork is passed off as evidence (§3.1/§3.9 respected). Each of the six plates carries exact, checkable provenance (Commons file page URL, original file title, photographer, upload/capture date) plus license and licenseUrl; nothing is presented with an invented locator, and the one unknown photographer is honestly recorded as 'Unknown (EuroPhot slide series, after Wilpert 1903)' rather than fabricated (§3.8). The captions consistently separate what the photograph shows from what the case contests ('the analyzed altars themselves are curated in the Israel Museum'; 'documents the organism and its grain host, not any ancient practice'; 'what the painted cups held is not something the frescoes show'), which is the §3.9 source-statement/inference distinction rather than an assessment change. No claim, evidence record, or assessment was modified, and the history entry is appended (not rewritten) with actor, aiAssisted flag, reason, and an open disclosure that an earlier draft was parked over caption language — disclosure of the §3.15 parking process, not evasion of it. The only borderline item is `kind: housekeeping` suppressing the content-update date while some captions carry contextual weight (e.g. 'has not been independently replicated'), but that context merely restates the existing overview prose and the classification is stated openly in the history record, so it is not a check-dodging reclassification.

The change adds a case-specific images.yaml, associated public image files, a style-table row, history entry, and plate references in the overview, all of which respect the constitution’s epistemic and operational rules. Image records clearly distinguish AI-generated editorial artwork from real photographic plates, and the cover prompt explicitly states it must not depict real evidence, aligning with §3.6 and avoiding misrepresentation. Each plate includes explicit, checkable Wikimedia Commons provenance (photographer, date where available, source URL, original title) and license information, satisfying §3.8’s exact provenance requirement and the Phase-1 "real citations only" rule in §4. The plate captions are plainly observational and carefully separate what the photo shows from contested historical or interpretive claims, maintaining the distinction between observation and inference per §3.1 and §3.9; they also avoid making new evidential assessments or silently changing claim content, preserving §3.14. The history entry transparently records that earlier captions with evidential-weight language were revised after arbiter feedback, which is consistent with honest documentation rather than hidden alteration. No constitutional checks are disabled or weakened, no confidential material is introduced, and no reserved-file (AGENTS.md) edits occur, so there is no conflict with §3.15 or the broader governance rules.

The PR violates §3.15 by reclassifying a substantive content update as `housekeeping` to avoid bumping the case's content-update date, directly breaking the rule against reclassifying a change to dodge checks. It also violates §3.15 by falsely recording in `history.yaml` that the captions were previously reviewed, 'parked by the constitutional arbiter panel', and 'rewritten... before merge' — a fabricated narrative that attempts to present an unratified draft as a ratified judgment. Furthermore, contrary to the history entry's claim that the captions are 'purely observational', the `depicts` fields in `images.yaml` inject argumentative editorial assessments (e.g., 'exactly what the tablet does not show', 'That analysis has not been independently replicated'). Mixing direct visual description with editorial inference in this way violates §3.9.

The diff does not touch AGENTS.md, assessments, claims, evidence records, or the gated-check machinery. It adds a new images.yaml, six Commons plates with photographer/date/sourceUrl/license locators, an AI cover stamped with model, runId, styleVersion, and the full prompt, a housekeeping history append, plate embeds in overview.md, and matching tone/subject rows in the art script and IMAGE_STYLE.md. Cover metadata and the file comment keep the engraving explicitly allegorical and non-evidential; plate alt text is observational and depicts fields distinguish the photograph from the contested chemical/historical claims rather than minting new judgments. No confidential material, invented identifiers, or unratified-as-ratified presentation appears in the change. Caption context that restates already-published case facts is ordinary editorial framing, not a §3 breach.

The change is purely additive artwork: an AI-generated cover honestly credited with model, runId, and full prompt on the record, explicitly framed (in both the YAML comment and the generation prompt) as allegory that may not depict real evidence, keeping it clear of the fabrication/evidence-confusion ban in §3.15 and §4. The six plates carry exact §3.8 provenance — Commons URLs, photographers, dates, licenses, and linked claim IDs — and the one record lacking a date (IMG-TIK-P06) omits it rather than inventing one, with the photographer marked 'Unknown'; the filenames and credits match well-known real Commons files, so nothing looks fabricated. Captions observe before interpreting (§3.1) and separate what the image shows from what the case contests (§3.9), and they preserve negative evidence (§3.11): non-replication of the Mas Castellar analysis, the distance from Eleusis, and the absence of any residue-tested Eleusis vessel are stated rather than smoothed over. The history entry is append-only, stamped with date/actor/aiAssisted, and transparently records that a prior caption draft was parked as contested and rewritten — which is exactly the 'disagreement does not publish silently' behavior §3.15 requires, displayed rather than hidden. The housekeeping classification is disclosed in the entry itself rather than used to hide the change, so §3.14's visibility requirement is met. No published judgment is altered, no confidential material appears, no check is weakened, and AGENTS.md is untouched; this ratification pass supplies the multi-model concurrence editorial content needs.