Recursive Evaluation Collapse: A minimal model for the multi-cycle dynamics of reflexive contamination

The third note in this series treated reflexive contamination as a static aggregate phenomenon and deferred its dynamics: the rate at which a categorical benchmark's discriminative function erodes per cycle of recursive use, the order in which the diagnostic markers assemble, and the limit toward which the process runs. The present note supplies the deferred treatment in minimal form. It introduces a bookkeeping model - a population updated partly as a function of the benchmark's own judgments - whose role is the role formal sources have played throughout this series: to mark the shape of a constraint, not to derive it. Within the model, the erosion of discriminative content under recursive use acquires a clock, and three trajectory regimes separate: a buffered regime in which exogenous influx slows erosion toward linearity; a self-feeding regime, engaged when evaluated populations supply the next cycle's training material, in which erosion compounds; and a cliff regime, entered when within-category signal falls beneath the judge's discrimination floor and judgments collapse onto surface features. The limit of the process is characterized: a population concentrated on judge-compatible features, over which the benchmark continues to issue well-formed judgments that report family membership rather than the task property. The three markers of the second and third notes are re-derived as co-movements of model quantities, with a predicted order of assembly that distinguishes contamination from genuine convergence. Seven falsifiable predictions follow, and a retrospective estimation sketch states what data would bear on them. The model is not fitted here; empirical execution is separate work. Domain instances beyond AI are deferred to the sixth note, synthesis to the seventh.