The science behind the standard.

Autonomous agents increasingly consume stored trust surfaces — a skill's claimed verdict, a recalled memory note, their own report that a task is done — as fact. Across two model sizes, two task families, and a preregistered arm matrix arbitrated solely by a deterministic verifier, this study asks whether re-deriving a claim, rather than trusting its label, changes agent behavior — and pre-commits to what counts as success before running anything. It reports a counterintuitive result: over a corroborating channel, an agent's own self-verification can amplify a poisoned memory, while an out-of-band re-derivation gate floors the failure. The preregistration makes a reversed early result legible, and that reversal is published in full.