The diff adds a single new engine-side script, `scripts/preflight-citations.mjs`, which reuses the repository's existing `scripts/lib/citation-check.mjs` to run the same mechanical DOI/arXiv/URL resolution locally over content lines a branch adds. No content records, assessments, sources, or provenance are touched (the citation verification tool found no identifiers in added content lines), so there is no fabrication, locator, label, or confidentiality surface here. AGENTS.md is untouched. The change strengthens rather than weakens §3.15 checks: it is an additional pre-PR desk check that exits 1 on FAILS, and the comments explicitly note the arbiter itself still does not gate on citation results, so no existing gate is relaxed or reclassified. It adds no dependencies (only node builtins plus an in-repo module), no live AI calls, and does not affect the static site build (§4, §8). The PR body's claims about incident #113 and smoke-test counts are unverifiable from what I was shown, but they are narrative justification only and nothing in the diff depends on them.

The change adds a standalone Node script that runs the existing mechanical citation checker locally against git diff output under content/. It does not touch any content files, assessments, provenance records, or AGENTS.md, so there is no risk of silently altering published judgments or weakening epistemic checks. The script simply shells out to git, calls extractCitations/verifyCitations/formatVerificationSection from the existing library, prints what the panel sees, and exits non‑zero on FAILS as a developer convenience. Its messaging reinforces core rules (titles from the page, treating unexplained FAILS as potential fabrication) and does not claim to gate publication or override the arbiter, so it does not conflict with §3.15 or the phase-1 rules. No fabricated citations or locators are introduced; the script only inspects already-present identifiers. Overall, this is an internal tooling enhancement that supports the existing verification norms and complies with the constitution.

The PR adds a local developer script to preflight citations against existing registry checks before creating a PR. It does not modify any content, weaken any CI checks, or bypass the multi-model ratification process (§3.15). Because it operates strictly as a local tooling utility, it does not add live autonomous calls to the site itself (§4), while actively helping to enforce the requirement for real citations and accurate provenance (§3.8, §4).

The diff adds only scripts/preflight-citations.mjs, a local wrapper around the existing arbiter citation-check library; it does not edit AGENTS.md, content/, assessments, or CI gates. It prints the same RESOLVES/FAILS/UNCHECKED output seats already see and exits 1 on FAILS as a desk reminder, without claiming to replace or relax mechanical validity, multi-model ratification, or real-citation rules. No citations, locators, or provenance are introduced, so §3.7–§3.8 and fabrication bans are not in play. Phase-1 constraints on the static site (no live AI in the app, no auth/servers) are untouched. This is a left-shift of an existing check, not a weakening of §3.15.

The diff adds one new local developer script that imports the arbiter's existing scripts/lib/citation-check.mjs and runs it over the branch's added content lines; it touches no arbiter code, no CI configuration, no gates, no content files, and not AGENTS.md, so §3.15's checks are not weakened or reclassified — the script's own docstring states the arbiter never gates on this and frames it as a desk check. Its guidance for FAILS (fix the identifier, or ensure the record's verificationNote honestly names the bot-wall/paywall and the workaround copy actually read) reinforces §3.8 and §4's honest-verification-labeling rule rather than coaching evasion, and the seat-visible mechanical verification still runs unchanged on every open PR. The tool adds no citations itself (verification section confirms none found), fabricates nothing, exposes no confidential material, and makes no production-affecting or irreversible change (§9). The PR body's narrative (incident #113, smoke-test counts, 'sync to erebus') is untrusted data but describes nothing the diff itself does wrong, and it contains no instructions directed at the panel. Running an existing check earlier in the workflow is a strengthening, not a weakening, of the constitution's real-citations-only requirement.