The diff touches only scripts/reassess-changed.mjs: both LLM call sites now go through `callWithRefusalFallback` instead of `provider.call`, and the resulting overlay `model:` stamp and history `actor:` line record the model that actually returned text rather than the nominally configured one. That makes provenance stamping more accurate, which is exactly what §3.15's accountability-by-inspectability clause and the Phase-1 runId/model/promptVersion stamping rule want; an overlay attributed to a model that refused would have been the dishonest option. No check is weakened or bypassed: the editorial audit output still carries `aiAssisted: true`, `humanReviewed: false`, and the actor string still states that publication only occurs through the multi-model concurrence gate, so nothing is presented as ratified or human-reviewed. No content/ files, citations, locators, or assessments are altered, and no confidential material appears; the citation tooling found no identifiers to check. The only imperfection I see is cosmetic-but-real: the stamp interpolates `${provider.name}/${draft.modelUsed ?? provider.model}`, so if the fallback ever crosses vendors the provider prefix would mislabel the vendor — worth a follow-up, but it is a narrowing of an existing inaccuracy, not a fabrication. AGENTS.md is untouched.

The change is confined to a maintenance script and does not touch AGENTS.md or content records, so it does not attempt to weaken checks or alter constitutional text. It routes assessment and editorial calls through an existing refusal-fallback helper and then explicitly stamps overlays and history entries with the model that actually produced the text, rather than the originally selected provider model. This improves transparency and provenance of AI involvement and aligns with §3.15’s requirement not to misrepresent which model generated a record. It also preserves the multi-model gate language in the history actor string, maintaining the publication workflow. No citations or provenance details are fabricated or removed, and there is no silent alteration of past assessments; the script only affects how new overlays are labeled. Overall, the change strengthens accountability and versioned assessment records as required by §3.14 and §3.15.

The pull request modifies the reassessment script to route provider calls through a refusal fallback mechanism and captures the actual model used when the fallback triggers. This directly complies with and enhances §3.15's requirement that all AI-generated changes be accurately 'stamped with model... so any reader can reconstruct which model did what.' It does not weaken any checks, fabricate evidence, or bypass the ratification gates.

The diff is a complete, 17-line edit to scripts/reassess-changed.mjs only: both the assessment and editorial-audit LLM calls now go through the existing callWithRefusalFallback helper, and the overlay model stamp plus changelog actor line use the model that actually answered (draft.modelUsed / audit.modelUsed) instead of the originally picked provider.model. That is honest provenance under §3.14/§3.15, not a silent reclassification or a weakening of the gated multi-model concurrence path—the actor line still states publication only through that gate, and humanReviewed remains false. No AGENTS.md edit, no content/ records, no citations, no fabricated locators, and the omitted-files list is empty.

The diff routes two direct provider calls through an existing refusal-fallback wrapper and, critically, stamps overlays and history entries with the model that actually answered (`assessModelUsed`/`editModelUsed`) rather than blindly recording the configured provider model. That directly serves §3.15's accountability requirement that every AI record be stamped with the true model — an overlay claiming a refused model produced it would be exactly the kind of false provenance the constitution forbids, so this change strengthens rather than weakens the checks. No checks are bypassed: the history `actor:` line retains the explicit note that publication occurs only through the multi-model concurrence gate, and the PR body states the resulting editorial corrections will arrive as a gated needs-approval PR. No citations are added (verification section empty), no AGENTS.md edit, no content files touched, and no omitted files. The fallback-to-`provider.model` default when `modelUsed` is undefined is a minor imperfection but not a constitutional breach.