The Right-Sized Mind: Model Scale and Behavioral Realism in Simulated Deliberation

This study was run and written up as Study 2 of the combined paper "The Inner Loop and the Ladder." Across nine models from 0.8B to 397B parameters, solo drift toward the textbook answer rose with scale, although the nine-point series was not monotonic. Generated-deliberation error was statistically tied from 4B upward. In a single-run split-model experiment, member fidelity and change dynamics tracked the thinking-model assignment more strongly than the speaking-model assignment. A neutral task sheet increased solo textbook-answer drift, an effect that survived four wordings and placements.