The diff touches only tooling (scripts/lib/citation-check.mjs, scripts/lib/arbiter-core.mjs) and their tests; no content/ records, no assessments, no AGENTS.md edit, and the verification section reports no citation identifiers added. The citation-check change makes extraction more faithful — `)` is no longer an unconditional terminator, and trimTrailing drops a trailing paren only when closing parens outnumber opening ones — so genuine Elsevier/ASCE DOIs are passed whole to the registry while prose and markdown-link parens are still trimmed; a fabricated identifier still FAILS, so the check is sharpened rather than weakened (§3.15). The capDiff change preserves the stated contract: sections are still selected by scrutiny tier, unspent reserve flows back in tier order on the second pass, and omitted files are still reported loudly, so nothing is silently hidden from seats and the fail-safe 'unsure' path is intact. I traced both new tests against the new algorithm (900/1000 and 300/300/100/500 at 900) and the reserve arithmetic produces the asserted kept/omitted sets. The one substantive tradeoff — a very large tier-0 change can now lose sections it would previously have kept — does not defeat the checks, since those losses appear in the loud OMITTED list and drive an 'unsure' vote rather than a silent pass. Claims in the PR body about a 'live incident' are untrusted narrative and did not affect my vote; the code alone supports the compliance finding.

The diff only changes internal arbiter tooling and tests; it does not touch content under `content/` or AGENTS.md, nor does it alter any epistemic or governance rules. The citation checker change makes DOI extraction more accurate by preserving balanced parentheses and trimming only clearly extraneous punctuation, which strengthens compliance with the real-citations and non-fabrication requirements in §3.7–§3.8 and §3.15 rather than weakening them. The diff-tier budgeting change in `capDiff` merely redistributes which diff sections are included under a fixed character budget and keeps omissions explicitly reported; it does not create any path where omitted material is silently treated as seen, so it preserves the integrity of the panel’s review process. No fabricated citations, locators, or provenance are introduced, and no existing assessments or records are silently altered. Because there are no omitted files and no signs of configuration that would hide violations elsewhere, there is no basis here for an "unsure" vote. Overall, the changes strengthen the mechanical checks instead of weakening them, so they comply with the constitution.

The diff correctly addresses bugs in the arbiter tooling without weakening the constitutional checks. Adjusting the DOI extraction to support balanced parentheses improves the accuracy of mechanical verification, preventing valid DOIs from being truncated and falsely flagged as failed. Allocating a reserved budget share for diff tiers prevents large unclassified files from starving the context window of content files, while retaining the strict requirement to report any omitted files. No rules regarding provenance, evidence, or check-weakening are violated.

The complete diff only touches arbiter tooling and its tests (citation-check.mjs, arbiter-core.mjs, two test files); AGENTS.md is untouched and no content/ records, assessments, or citations are published. The DOI extractor change stops amputating legal parentheses so identifiers like 10.1016/S0140-6736(20)30183-5 are resolved whole instead of FAILS-truncated, which tightens §3.7/§3.15 “real citations only” rather than weakening it; unbalanced prose/markdown parens are still trimmed. capDiff still reports every omitted path and still fills leftover budget in scrutiny order; the reserved 50/50 split only prevents unclassified tier-0 bulk from starving content/, which is the opposite of silent truncation or reclassification to dodge review. Tests pin both behaviors. Nothing in the shown change fabricates provenance, presents unratified judgment as ratified, or relaxes the gated checks.

This PR repairs two defects in the arbiter's own tooling without weakening any check. The citation-check fix stops truncating DOIs that legally contain balanced parentheses (Elsevier S-prefix, ASCE), so genuine citations are no longer misreported to seats as FAILS — that strengthens the real-citations-only rule (§3.8/§3.15) rather than relaxing it, since extraction still feeds the same resolve-and-report pipeline and nothing auto-passes. The capDiff change reserves half the packet budget for the content tier so an unrecognized tier-0 directory cannot push every content/ file into the omitted list; crucially it leaves diffTier's classification untouched, keeps the 400k budget, and preserves the loud-omission contract ({ text, omitted }) that forces 'unsure' votes and parking rather than silent publication, so the fail-closed property of §3.15 is intact. No reclassification of any path occurs, AGENTS.md is not edited, no content/ citations are added (the verification section confirms none, and the DOIs in the new tests are fixtures in test code, not content records), and no confidential material appears. The added regression tests directly assert both the balanced-paren extraction and the no-starvation guarantee. Ordinary tooling hardening, properly parked before this panel for ratification.