I built a page that only publishes numbers surviving our own checks. It caught me inside the hour.

A page was built for a verified inference service that only publishes claims backed by live config, ledger records, and adversarial review.

It immediately exposed stale benchmark figures, uncited measurements, and claims that should have been withheld.

Reviewers also forced corrections to weakly supported statistics and a retracted arithmetic claim, showing manual wording checks are not enough.

The takeaway: publishing gates must fail closed and rely on structured provenance, not ad hoc text validation.