Identifiability of sensory adaptation and readout plasticity in simulated visual behavior

Sensory adaptation and readout learning can produce similar visual improvements. We ask which measurements separate their predictions and when attribution should remain uncertain in an image-driven observer. Twenty photographs drove a shared plastic-normalization population with separate discrimination and appearance readouts. Sensory and readout states matched at three development accuracies but differed under held-out noise and transfer conditions. After projecting out criterion and noise directions, additional measures increased the local mechanism angle from 5.77° to 85.33°. Flexible report noise or appearance weighting weakened appearance-curve constraints. Initial training reduced further readout gains, and shared sensory noise reduced modeled sensory benefit. Calibrated compatibility sets retained ambiguity at weak effects. At an uncalibrated sensory rate, nuisance shifts produced up to 14.45% wrong singletons with 1,728 responses. Known-parameter comparisons and exact-forward fits separated observation limits from candidate-support and scoring effects. Under sequential acquisition from a dynamic joint generator, frozen-state fitting selected the correct family but produced 0.6854 relative initial-state error. History-aware fitting reduced that error to 0.0082, while an incorrect dynamic assumption caused errors in other settings. Public threshold summaries lacked the inputs for a matched prediction test. Attribution in this model therefore requires distinguishable predictions, adequate observation assumptions and an acquisition model. Local separability alone does not guarantee calibrated mechanism recovery.