A Dynamical-Systems Approach to Generative Witness Simulation for Legal Training

Deposition training requires attorneys to recognize and manage dynamic interpersonal behaviors that existing legal AI benchmarks do not measure. We introduce Witness Sim, a generative deposition simulator and evaluate its behavioral realism using emotion-vector analysis. We project hidden activations onto 23 directional emotion vectors to compare emotional trajectories between real deposition transcripts and synthetic ones generated across ten configurable witness archetypes. Synthetic depositions reproduce statistically significant temporal emotional structure but are systematically compressed in emotional intensity relative to real ones. Event-level analysis shows that combative witness behavior reflects a persistent dispositional mode rather than a reactive event, while transient personality shifts produce recoverable spikes in residual emotion space. These results suggest that emotion-vector trajectories provide a principled realism metric for LLM-based legal simulation and identify concrete calibration gaps between simulated and real deposition dynamics.