Reciprocal AI–Neuroscience Translation: From Biological Mechanism to Engineered Failure, Rescue, and Biological Retest

This preprint introduces reciprocal AI–neuroscience translation as a typed experimental protocol. A biological relation is mapped into an engineered analogue; capacity-matched models are tested under held-out lesion, drift, compensation, and rescue; the resulting discriminating signature is then returned as a predeclared biological retest. The paper reconstructs the SAN genealogy of this reciprocal method, distinguishes it from generic biomimicry and predictive fit, reports a deterministic reference application spanning 36 paired synthetic systems, seven intervention families, four models, and 1,008 model-intervention evaluations, and supplies a bounded Lean 4 verification layer. The correctly typed model achieved intervention-effect RMSE 0.01454, versus 0.37497 for endpoint baselines and 0.46317 for a deliberately wrong typed map. These results validate only the declared synthetic and formal models; biological confirmation requires a successful, discriminating return experiment. ## Evidence boundary The application is deterministic, synthetic, and deliberately small. Lean checks only the declared bounded logical leaves. Neither layer establishes that the proposed translation map is a biological mechanism, and neither substitutes for a held-out biological intervention. The manuscript is a preprint and has not been peer reviewed.