Four target models share an initialization and fixed scoring metric. A and B use the predeclared20 exposed categories; C uses all50 for actual validator behavior. D reports a difference of retrieval effects, not mediation. Intervals resample facts10000 times with seed20261118 and condition on these training trajectories. A has point estimates; paired intervals belong to B–D. Input matching uses its separately fixed input metric. The comparison does not establish historical training influence. All11 fixed readouts and 43 matched baselines for all4 targets appear in the accompanying CSV.
