Self-confirming blind regions in adaptive screening: an audit-rate bifurcation with critical slowing down hidden from the monitored proxy

Adaptive quality screening — a learned model that decides which units to verify, retrains on what it verifies, and filters production by its own predictions — is a closed loop in which the observation channel is generated by the system’s own state. We give a low-dimensional dynamical account of this loop. The fraction of confidently misclassified units (a self-confirmed blind region) obeys, in mean field, a contact-process law whose control parameter is the audit rate. Under simple contagion the loop undergoes a transcritical bifurcation; under cooperative contagion the bifurcation becomes a fold, producing bistability and hysteresis in which the blind region is irreversible by audit rate alone. The mean-field law is derived as the limit of an agent-based contact process and confirmed numerically. The bifurcating coordinate is unobserved: critical slowing down is present on approach to the threshold yet absent from the monitored proxy, so early-warning failure here is observational — the critical mode lies in the null space of the deployed measurement — rather than a matter of the order of the transition. This is a structural/possibility result, motivated by but not validated on potency screening of cell-derived therapeutics.