When Correct Biological Priors Backfire: Auditing Reward Priors for Agents that Transfer Across Brain Tissue

ABSTRACT: Agents that act in closed-loop biology, such as autonomous microscopes, patch-clamp robots, and experiment planners, reuse policies across specimens and are routinely given domain knowledge as reward: ``sample the tissue interior'' or ``stay where the target cells are''. We show that this combination can fail silently, even when the knowledge is correct. BrainNav converts three public mouse-brain spatial transcriptomics datasets (MERFISH hypothalamus, Slide-seqV2 hippocampus, and Visium coronal section) into 14 tissue-shaped navigation tasks whose goal is oligodendrocyte-rich white matter. A tissue bridge embeds every section into one code space, and warm-starting a tabular policy through it raises greedy success in all eight section and cross-platform pairs (gain $+0.14$ to $+0.31$, $p<0.001$, 154 of 160 seed-pairs). Adding a state-based prior at a typical weight cuts transfer success to 0.02--0.04 in every pair. A correct cell-type atlas fails the same way, and agents cycle in place until time-out. We prove a critical weight for any bounded state-based prior. We also show that the time-blindness introduced by the transfer bridge is what turns bounded goal deferral into goal abandonment: the safe weight is horizon-independent for a time-aware agent but shrinks roughly as $1/H$ for a time-blind one. An exact dynamic-programming audit, which takes seconds per section and needs no training, predicts the learned collapse to within 25\% in 19 of 23 section--horizon settings. Rewriting the same priors as potential differences keeps transfer success at 0.49--0.60 (unshaped: 0.59) at every tested weight up to $w=1$. Together these give a simple audit-then-shape protocol for agents that carry biological priors across specimens.