Researchers release PAWS dataset for policy-driven financial agent simulation

Quality checks on 2,522 stratified action samples showed 89.4% initial agreement between independent AI and human reviewers on interaction mode, with disagreements adjudicated afterward. Case studies on the 2008 short-selling ban and 2001 decimalization demonstrate the dataset's ability to recover documented policy timelines and market patterns. A replay study also found that high overall accuracy can mask failures to detect rare stakeholder actions, highlighting action timing and calibration as key challenges. The authors position PAWS as an auditable substrate for evaluating agent influence, policy-response cascades, and action-outcome alignment in historically grounded financial simulations.