The diff adds no citations (verification tooling found no identifiers), touches no content/ records, and does not edit AGENTS.md. The main functional change, `survivingObjections`, surfaces dissenting cross-model checks on ratified case pages with a link to /panel for full reasoning — this strengthens rather than weakens §3.15's 'disagreement does not publish silently' and §3.11's negative-evidence norm; the accompanying test pins the count to exactly the dissenters the ratification tolerated. The prompt changes bump PROMPT_VERSION honestly (v1→v2 with an inline note in all three scripts, preserving the runId/promptVersion stamping required by §3.14/§3.15) and add a sensitivity requirement that makes verdicts state their single load-bearing evidence record — a tightening of calibrated uncertainty (§3.13), not a dodge. The method-page edits restate existing policy (independence groups per §3.10, honest verification labels per §4) in prose; the 'literature-watch queries are public, versioned files' sentence is a self-description I cannot verify from this diff, but it is a claim about repository practice rather than a fabricated citation or locator, and nothing here removes provenance or relabels a draft as ratified. Minor nit: the seat label strips ', independent' from the model string for display, but the full run records and /panel retain the provenance, so no §3.8 concern.

The diff does not touch AGENTS.md and introduces no citations, so there is no risk of fabricated or unverifiable sources under §3.7–§3.8. The new `survivingObjections` function and its UI usage simply expose existing dissenting cross‑model checks for ratified cases, making tolerated disagreement visible rather than hiding it; this strengthens transparency and aligns with §3.3, §3.11, and §3.14. The method-page edits clarify that repeated reports are not independent evidence and that corpus assembly is inspectable, which operationalizes §3.10 and reinforces the existing non-fabrication and inspectability requirements in §3.15. The prompt updates (v2) add a sensitivity/single-thread requirement, making assessors state which evidence record is load-bearing and whether the verdict survives its removal, which clarifies evidential dependence and supports calibrated uncertainty under §3.13 without weakening any checks. No append-only or ratification constraints are relaxed; processes for multi-model concurrence remain intact, and the changes are additive and more conservative. Overall, the change tightens epistemic discipline and public explanation of existing policies rather than undermining them.

The changes directly support the constitution's transparency and epistemic requirements. Surfacing 'surviving objections' on the case page ensures that tolerated multi-model dissent is visible rather than sanitized away, adhering to §3.15 (inspectability) and §7 (interface symmetry). Adding the 'single-thread test' (sensitivity) to the prompts strengthens calibrated uncertainty (§3.13) by forcing models to identify fragile verdicts. The additions to the method page accurately document existing requirements regarding independence (§3.10) and accountability (§3.15).

The diff adds transparency rather than weakening §3.15 checks: ratified case pages now surface tolerated dissent from existing cross-model checks (verdict + first synthesis sentence + link to /panel), which matches the constitution’s requirement that disagreement not be hidden. Prompt bumps to check/assess/reconsider v2 require naming the single most load-bearing evidence record and whether the verdict survives without it; that is calibrated uncertainty (§3.13), not a dodge of ratification, and overlay assessments remain append-only. Method-page copy restates independence groups (§3.10) and inspectable corpus assembly; no AGENTS.md edit, no citations/locators added or fabricated, no published judgments mutated, and omitted-files is empty. Tests pin that objection count equals tolerated dissenters on ratified cases. Ordinary UI/copy judgment, not a rule break.

The diff strengthens rather than weakens the §3.15 gates: all three prompt edits (cross-model-check, reassess-changed, reconcile-contested) add a mandatory sensitivity line — name the single evidence record the verdict most depends on — and honestly bump prompt versions v1→v2 with explanatory comments, with no change to schemas, RATIFICATION_MIN_PANEL, or CI. The new survivingObjections function is read-only over append-only records and implements the constitution's requirement that disagreement not publish silently: ratified cases now display their tolerated dissenters' verdicts with a link to /panel for full reasoning, which is more transparency, not less. Stripping the word 'independent' from the seat label is display formatting while the full record remains linked — not provenance removal. The method-page additions restate §3.10 (repetition is not corroboration, independence groups) and describe an inspectable, versioned literature-watch corpus; nothing in the diff contradicts those claims, so there is no positive evidence of fabrication. A 21-line test covers the new domain function per §8, no citations were added so the empty verification output is unremarkable, and AGENTS.md is untouched.