PaperTrail

Insurers deny medical claims over payer rules nobody writes down. We built a self-improving harness that learns those rules from the denials. A scrubber agent (Claude Sonnet) predicts whether each claim will be denied. A watcher agent (Claude Opus) reads the insurer’s denial codes and writes new executable rules. A code gate replays every proposed rule against all past outcomes and promotes it only if it fixes a real miss with zero regressions, so no model grades its own work. Rules are versioned in MongoDB Atlas, and the harness also merges its own rules into more general ones. On 22 held-out claims it never trained on, denials caught went from 5 of 11 to 10 of 11 (balanced accuracy 0.636 to 0.818) over 3 learning rounds. The gate rejected 10 of 26 proposed rules.