Developer builds self-auditing claims page, catches his own unpublished data within an hour

A developer running a verified AI inference service built an auto-generated transparency page designed to publish only numbers that pass internal quality checks. Within an hour of launch, the system flagged two results the company's own ledger had explicitly marked as not cleared for external publication. Further investigation revealed additional figures with no ledger entries at all, meaning they had been sourced from informal internal handoffs rather than verified data. An adversarial review of a remaining claim uncovered three specific defects, including a statistic presented as a stable range that was actually drawn from two small, underpowered subgroups, and a cited interval that had no supporting sample size or date in the ledger. The episode led the developer to adopt a fail-closed publishing rule, where a figure can only go live if its ledger entry exists and carries no hold marker.