This is a tooling/eligibility change (scripts/lib/reconcile-core.mjs, reconcile-contested.mjs, a vitest file, one appended DECISIONS.md line) with no content/ records and no citations, so the citation and provenance rules are not engaged. AGENTS.md is untouched, and DECISIONS.md is only appended to, consistent with its append-only header. The change broadens, rather than narrows, when the house must answer dissent: cases where the panel splits on a load-bearing claim under a unanimous case verdict now become eligible for a reconsideration that must engage each split claim by id, and the standoff guard (draft.promptVersion === PROMPT_VERSION) is left intact so no re-litigation loop is created. Nothing here relaxes ratification: reconsideration overlays remain stamped as AI drafts ('Append-only; not reviewed') and the previously ratified `reconciles` rule still forces a fresh blind check before standing can rise, so this does not let a case self-ratify by converging on its judges (§3.15). The duplicated graded-scale/strict-majority logic in reconcile-core.mjs cannot be diffed against load.ts here, but its behavior is documented, tested (including the exact 4-of-5 open-verdict shape and the within-one-step non-split), and any divergence would only mis-trigger engagement rather than weaken a publication gate. The 'live dry run' described in the PR body wrote nothing and is not part of the diff, so no unratified judgment is presented as ratified.

The change does not touch AGENTS.md and instead adjusts internal automation around reconciliation eligibility and dissent handling, plus updates DECISIONS.md to document the decision — consistent with §3.14’s versioning requirement. The new reconcile-core module and its tests explicitly mirror existing ratification logic in load.ts, tightening, rather than weakening, the conditions under which reconciliation is triggered and making claim-level dissent explicit; this strengthens the multi-model check process and does not reclassify changes to dodge checks (§3.15). The reconciliation script continues to label overlays as reconsiderations and, per the referenced prior decision, leaves ratification to independent blind checks, maintaining the separation between reconsideration and final standing. No sources, citations, or provenance are added or altered, so there is no risk of fabricated records or dishonest verification (§3.8, §3.15). The logic is append-only in behavior (it broadens what is considered contested but does not silently alter past judgments) and improves transparency by logging the dispute shape, aligning with the constitution’s emphasis on inspectability and explicit disagreement handling.

The change strengthens the dispute-resolution system by ensuring that splits on load-bearing claims are actively engaged by the reconciliation loop, rather than deadlocking indefinitely. It updates the LLM prompt to require explicit engagement with claim-level dissents, upholding the requirement that assessment changes record what changed and why (§3.14). The logic exactly mirrors the existing ratification rules without weakening any multi-model concurrence requirements or bypassing checks (§3.15).

The diff only expands reconciliation eligibility so it matches the already-contested shapes (case-verdict dissent or a load-bearing claim split), extracts that rule into a tested helper, and appends a DECISIONS.md entry. It does not edit AGENTS.md, does not add citations or content records, and does not rewrite published assessments. The standoff guard and the existing reconciles/no-self-ratify path are left in place, so a house reply still cannot mint concurrence without a later blind panel. Prompt-version string reuse is sloppy inspectability, not a §3.15 check-weakening. Ordinary incompleteness relative to unseen load.ts is not a visible constitutional break.

The change extends the reconciliation loop's eligibility so that a panel split on a load-bearing claim under a unanimous case verdict triggers mandatory house engagement with the dissent — this strengthens, rather than weakens, the §3.15 requirement that disagreement not sit silently: the dissent packet now carries per-claim verdicts and seat reasoning, and the prompt requires answering each split claim by id. Critically, the change does not let reconciliation mint concurrence: the ratification gate is untouched, the standoff guard is preserved (extended to claim-level standoffs), and both the PR body and the appended DECISIONS.md entry explicitly rely on the prior day's fix that a `reconciles`-stamped draft must wait for a fresh blind panel. Assessments remain append-only overlays stamped with runId (§3.14), and DECISIONS.md is appended to, not rewritten. No content/ files are touched, no citations are added (verification found none, consistent with the diff), no AGENTS.md edit, no confidence material, no live AI calls added to the site. The one risk — reconcile-core.mjs mirroring ratification() logic from load.ts, which I cannot compare directly — is ordinary duplication risk the code itself flags ('load.ts is the authority'), and even a divergence would mis-trigger drafts that still cannot self-ratify, so it is not a constitutional violation.