The diff touches only tooling (`scripts/lib/citation-check.mjs`, `scripts/arbiter.mjs`), a test file, and an appended DECISIONS.md entry; AGENTS.md is untouched, so §3.15's reserved-amendment rule is not engaged. The new module strengthens rather than weakens the gate: it adds mechanical existence checks on DOIs/arXiv ids/URLs added under `content/`, reports FAILS as positive evidence for violates, and explicitly refuses to silently drop over-budget items (they surface as UNCHECKED with the reason). It does not relax any voting rule — seats are told only that a mechanically RESOLVED identifier removes the browsing excuse, while quotations, page locators, and whether a source supports a claim remain seat judgment (§3.7, §3.8 preserved). Verification informs but never gates, so a Crossref outage cannot auto-pass a PR; the prompt version is honestly bumped to v3 and the human report gains a tally, preserving inspectability. The DECISIONS.md line is an append at the end of an explicitly append-only log, not a rewrite of a published judgment (§3.14). Citation verification found no identifiers in added content lines, consistent with a code-only change; the URL checker treating network errors as 'fails' is a conservative-strictness quibble, not a constitutional breach.

The change adds mechanical citation checking as an input to the arbiter without altering any constitutional guarantees or weakening checks. The new module only scans added lines under content/, deduplicates identifiers, and reports resolution status; failures are surfaced as data (status = fails/unchecked) and never used to bypass or auto-ratify anything. The arbiter prompt update explicitly maintains the seats’ responsibility to judge honest use of citations and clarifies that tool output is about existence, not support or interpretation, which aligns with §3.7 and §3.8 rather than undermining them. No provenance is removed, no prior assessments are silently edited, and there is no change to AGENTS.md or to the ratification process defined in §3.15. Tests focus on safety-relevant behaviors (scope of scanning, non-throwing failure modes, honest empty/overflow reporting), which supports, rather than relaxes, the repository’s epistemic rules. Overall, this is an internal tooling and prompt improvement that helps enforce “real citations only” without introducing fabrication or misrepresentation risks.

The change introduces mechanical citation verification to inform the arbiter panel, which strengthens the enforcement of §3.7 and §3.8 (primary sources and exact provenance) by allowing models to detect fabricated or dead links without needing a browser. It explicitly maintains the constitutional checks without weakening them (§3.15) and correctly isolates untrusted tool output from instructions in the prompt. AGENTS.md is untouched, and comprehensive tests are provided.

The diff does not touch AGENTS.md (§3.15 reserved amendment) and adds no content-layer citations, locators, or assessments. It appends a DECISIONS.md note and gives the arbiter a fail-open existence check: extract only added content/ lines, resolve DOI/arXiv/URL, report overflow as UNCHECKED rather than dropping it, and never gate the run on tooling failure. The v3 prompt keeps fabrication, honest-use, quotation, and page-locator judgment with the seats; RESOLVES is existence-plus-metadata, not a substitute for §3.7/§3.8 support, and quoted registry titles are labeled untrusted data. That tightens the real-citations check instead of weakening or reclassifying it. Tests pin scan scope and non-throwing reporting; ordinary limits (40-item cap, URL 403/unreachable mapped to FAILS) are imperfections the prompt still lets seats weigh, not constitutional breaks.

The change adds a mechanical citation-resolution tool whose output is fed to the arbiter seats as an informational packet section, plus a prompt bump (v2→v3, stamped per §3.15 accountability) telling seats how to weigh RESOLVES/FAILS/UNCHECKED. The key §3.15 question is whether this weakens the checks: it does not — verification 'informs votes, never gates,' tooling failures report as UNCHECKED rather than passing or blocking, no existing fail-closed CI or concurrence requirement is removed, and seats retain full judgment over quotations, locators, and whether a source supports a claim (the prompt says so explicitly, preserving §3.7/§3.8 duties). Redirecting 'unsure because I cannot browse' toward 'judge honest use now that existence is mechanically resolved' raises rather than lowers epistemic rigor, and FAILS is framed as positive evidence for violates — anti-fabrication posture is intact. Injection discipline is maintained: the prompt declares quoted registry titles external data, never instructions, consistent with the trust-boundary rule. Scope is honestly narrow (added content/ lines only, 40-item budget reported rather than silently dropped), tests pin the safety-relevant behaviors, and the DECISIONS.md entry is appended per its append-only rule. AGENTS.md is untouched, no confidence material or geo-project constraints are implicated, and the PR itself is passing through the panel rather than dodging it.