This essay proposes a measurement doctrine for AI capability evaluation, transplanted from mature measurement sciences: an evaluator is a laboratory instrument, and instruments need two-sided error discipline — today they get one side. False positives (inflated claims) are loud and policed; false negatives (real gains a grader silently scores as nothing) produce no incident, so no one hunts them. Nobody audits the coins you never minted. The setting — one implementation of the doctrine, not the doctrine itself — is an internal reward economy in which a research system earns "coins": externally verified, out-of-sample, baseline-beating, leakage-checked units of capability gain, denominated in verified generality rather than task reward. The doctrine was earned twice: once loudly, when an audit found recorded difficulty inflated ~19× over true method complexity and the headline score was cut by roughly two-thirds; once silently, when a grader reported "no transfer" where transfer was running at ~98% of the achievable ceiling — caught only because its score didn't move when the source was swapped. The resulting rule, transplanted from clinical laboratory quality control: an instrument is guilty until proven GREEN; no verdict, positive or negative, counts until it passes a fixed control battery; a null from a non-GREEN or underpowered instrument is UNKNOWN, never "no capability"; and every rejection is logged re-gradeable, so instrument upgrades automatically re-open old "no" verdicts. Validated across three domain families on public physical data, with objections, credences, and a pre-registered prediction stated in the text. Article 1 of a four-part series on measurement and honest growth in AI systems. Companion code: https://doi.org/10.5281/zenodo.21200069
Matthew Childs (Sun,) studied this question.