Information Organization and Evidence Efficiency in Policy Induction

Professional interfaces organize information to support human observation and judgment, but whether this organization benefits foundation models requires evaluation. We compared two text representations of identical chronological state-action observations generated from public market data using study-defined hidden policies: flat tables (FLAT) and panels inspired by professional trading interfaces (PANELS). Observations were matched at each evidence budget, with execution and evaluation conditions held constant across representations within each model. We introduced Evidence-to-Criterion (E2C) to record the first tested observation budget at which an induced policy met a prespecified recovery criterion for action reproduction on separate evaluation states. On the evaluated task, FLAT showed significantly greater evidence efficiency in policy induction than PANELS for both models. These findings underscore the importance of information organization and motivate studying its role in model inputs and training data across domains.