What the Mind Does Not Know, the Eyes Do Not See: How Learned Representation Governs Surprise, Deployment and Change Detection

Abstract

An event can reach the eye, be transduced and remain decodable in sensory features without becoming a meaningful change for an agent. Calling this failure “inattention” hides a deeper dependency: prediction error is measured against a learned model, and a model cannot be selectively surprised along a distinction it has collapsed. We call the result model-relative blindness. In the first confirmatory experiment, represented and collapsed agents received identical observations, actions, noise and changed-world evidence; their training ecologies differed only in whether a cue predicted causal outcome. The cue remained almost perfectly decodable from early features. Represented agents preserved it in an action-facing state. Collapsed agents averaged the cue classes. A changed law placed at that pooled expectation produced near-perfect inverse ranking: ordinary events looked more surprising than the genuine change, and 90% recall required a 100% ordinary-event false-positive rate. A second preregistered experiment froze the trained models and moved the changed law away from the pooled expectation. It confirmed three regimes. Below the ordinary old-law error scale, detection remained inverted. At the predicted scale of 1.0, collapsed-model AUC crossed chance (.4864 and .4918; both 95% intervals included .5). Beyond that scale, the change became detectable as a generic anomaly, but an unchanged consequence gate remained near chance in balanced accuracy even at distance 2. Diagnostic experience, rather than threshold adjustment, rebuilt the missing causal distinction. Together the studies establish a theory-predicted crossover from inversion, through confusion, to anomaly detection under representational collapse. They also separate detecting that something is wrong from representing what is wrong. The result is computational; it does not establish human phenomenology, biological identity or machine consciousness.