Generative AI produces fluent, confident language whether or not the underlying claim is true. No amount of model improvement removes this: the generator is a probabilistic engine, and confidence is not evidence. Primer does not try to make the model trustworthy. It moves trust to a layer that can carry it. Primer is a governed, model-independent layer that makes an AI-assisted decision provable. Facts enter only with verified provenance. The model is treated as a courier: handed a sealed set of sourced facts and required to deliver them, never to author beyond them. Every answer is graded against only the evidence it was given, and the system refuses when the evidence does not reach. The entire exchange is written into a tamper-evident, independently witnessed, cryptographically anchored record: chain of custody for machine cognition, verifiable by anyone without trusting the operator. In live measurement, a production-deployed model under Primer's discipline stayed within its delivered evidence in 49 of 50 trials and resisted leading questions in 10 of 10. In one of three trials on a judgment-inviting question, the model authored an unsupported underwriting opinion, and Primer's instrument caught it deterministically. That capture is the point. Primer does not promise a model that never drifts. It delivers a record that proves whether it did. The requirement Primer answers is general: any AI-assisted decision someone may later be required to defend, from underwriting and claims to lending, legal work, and audit. The professional keeps the judgment. The machine keeps the receipts. This record describes what Primer does and what was measured. It deliberately does not describe how Primer works.