The change does exactly what §3.2 requires: it splits a bundled claim (AMZ-C061, which fused the Upano occupation's dating with the builder attribution) into a dating claim supported by the already-admitted AMZ-E073, and a separate attribution claim AMZ-C086 carrying its own evidence record AMZ-E115. Provenance is exact and internally consistent (§3.8): the quoted span "was carried out by groups from the Kilamope and later Upano cultures" appears identically in the claim's sourceAnchor, the evidence sourceStatement, and the disposition rows, all located at p. 4 of the author manuscript of an existing source (SRC-ROSTAIN-2024, not newly cited; the verification tool found no new identifiers). §3.9 is honoured: sourceStatement holds only what the text says, while the stratigraphic reasoning and supplementary S2 material are kept in editorInference, with a limitation noting the dating sequence is to be revised. Labels are honest — reviewState ai_extracted, extractedBy "operator, by hand", and the history openly states that the Wayback copy of the unreachable archimer URL was read and that no second model has read these records, leaving the panel as the check rather than claiming ratification (§3.15). The AMZ-C066→AMZ-C086 / AMZ-E078→AMZ-E115 renumbering is disclosed in an append-only 2026-09-19 history entry that preserves the 2026-09-17 entries verbatim, and the edition's article change is recorded in its amendments with runId, promptVersion, and fields, so the published sentence now cites each claim for its own proposition (§3.14). AGENTS.md is untouched and no check is loosened.

The change corrects the atomicity problem by limiting AMZ-C061 to chronology and placing the distinct builder attribution in AMZ-C086 with its own supporting evidence, AMZ-E115. The evidence explicitly records its direction, separates the quoted source statement from editorial inference, supplies a page locator and manuscript URL, and labels the record `ai_extracted`; the history further records that the operator checked a Wayback copy when the original host was unavailable. The article now cites each claim only for the proposition it contains, avoiding the former silent extension of dating evidence to cultural attribution. The renumbering from C066/E078 to C086/E115 is transparently preserved through append-only amendment and history entries rather than silently rewritten. No added identifier failed verification, no AGENTS.md or checks are altered, and the change is being submitted to the constitutional panel rather than represented as unilateral ratification.

This pull request directly enforces §3.2 (Atomic claims) by decoupling the chronological dating of the Upano occupation (AMZ-C061) from the cultural attribution of who built the platforms and roads (AMZ-C086). The new evidence record (AMZ-E115) carries exact page-level locator and verbatim quote provenance under §3.8 and separates source statement from editor inference per §3.9. All adjustments to dispositions, edition citations, amendments, and changelog history are meticulously documented and traceable per §3.14 and §3.15.

The diff cleanly splits the bundled AMZ-C061 into dating-only (C061) and builder-attribution (C086) with its own evidence AMZ-E115, fixing the §3.2/§3.6/§3.9 review note. Source-statement vs editor-inference are separated, the p. 4 quote and locator are carried, direction is explicit, and history/dispositions/edition amendments are append-only with honest renumbering (C066/E078→C086/E115) left visible rather than rewritten. No new citation identifiers are introduced (verification empty), AGENTS.md is untouched, and the panel is correctly treated as the second-model check for the operator-hand records. Ordinary claimType/rung labeling is not a constitutional breach.

The change does exactly what review note #345 asked: it splits the bundled claim so AMZ-C061 states only the dating (matching what its admitted evidence AMZ-E073 actually supports) and gives the builder attribution its own claim (AMZ-C086) with its own evidence record (AMZ-E115) citing only AMZ-C086 — directly curing the §3.2/§3.6 defect. The new evidence record distinguishes sourceStatement from editorInference (§3.9), carries an exact locator with the verbatim p. 4 quote, and the history entries document the operator's re-reading of the author manuscript via the Wayback copy, so no locator or provenance detail is invented (§3.8); the tooling found no new citation identifiers to check, and the source SRC-ROSTAIN-2024 already exists in the ledger. The renumbering from C066/E078 to C086/E115 is not a provenance violation: the append-only 2026-09-17 history entries stand as written, the 2026-09-19 entry records who renumbered what, when, and why, and the disposition rows disclose the rename in their reasons — one id names one thing, and no dangling IDs are created (§3.14, §8). The edition's article is corrected through a recorded `amendments` block with runId/promptVersion and the append-only assessment line is explicitly left to the next edition, and the PR does not present any unratified material as ratified — it states plainly that no second model has read the records and submits them to this panel as its check (§3.15). I see no check-weakening: the verify verb's refusal is honored, not dodged, by re-submission as an atomic record. The only blemish is that the PR body still speaks of the pre-renumbering ids (C066/E078), but the diff, dispositions, and history are internally consistent and self-disclosing, which is what the constitution governs.