Structured Clinical Memory as Context for Probabilistic Decision Models: A Synthetic Evaluation with Jev
Summary
- Researchers studied whether compact structured memory improves probabilistic decision models using 30 synthetic clinical cases and 120 typed questions with Jev.
- The raw-vignette arm achieved 90.0% accuracy, while adding structured memory achieved 93.3% accuracy.
- Choice log loss decreased from 0.258 to 0.175 and the binary Brier score improved from 0.054 to 0.022.
- Structured memory added an average of 310 input tokens per call, while observed mean latency remained unchanged at approximately 0.5 seconds.
Why it matters
Evaluating structured memory as context for decision models demonstrates potential improvements in accuracy and calibration without increasing latency, providing a reusable state interface for AI applications.