Public works wants models blocked from promotion unless evaluation is complete. What CI practice fits?

Public-works release discipline treats comprehensive evaluation as a hard CI promotion gate: the build must not advance until the full eval suite—including cross-platform checks—passes. That practice keeps incomplete or host-skewed models out of production channels that affect streets, permits, and resident services. Allowing promotion when a single offline BLEU peek looks fine substitutes one proxy for the suite and can miss toxicity, factual, or cross-host regressions the rest of the suite would catch. Skipping eval gates for Friday afternoon hotfixes institutionalizes risk at the worst time for on-call coverage and often becomes the permanent exception path. Promoting first and running the suite only if residents complain inverts the control: harm discovery shifts from pre-release measurement to public incident response. Exam caveat: gates exist to fail closed; partial peeks and post-complaint evaluation are not substitutes for a completed suite. Operational check: wire the municipal eval suite into CI so the promotion job waits on all required checks (quality, safety, cross-platform), deny override without documented risk acceptance, and retain green/red artifacts with the release record for auditors.