Preserve the evidence. Generate missing responses. Inspect the reward.
17 + 14 turns · Blinded models · Turn-by-turn review
NEJM · 2009 · 361:511–518
History, laboratory results, neck imaging and tissue findings form the evidence of the environment.
16 history facts · 44 findings · Source-page links
Read the report and original figures ↗Preserve reported evidence; generate responses for undocumented requests.
Request → Evidence match → Generate missing items → Check / retry → Cache / ledger
Case story · Documented history · Known results · Requested item
Diagnosis / overlap checks · Up to 2 attempts · Cached within each episode
Loading recorded runs…
Generated values are not observed patient measurements. Current checks look for diagnosis terms and documented-item overlap; clinical consistency still needs review.
Diagnosis · Encounter cost · Unsupported claims and errors
R = wdx × (score − 1) / 4 − wcost × min(1, cost / scale) − Σ penalties
Simulator-generated results and unsupported agent claims are reviewed separately. A’s penalty concerns unsupported claims in its final rationale.
Illustrative reweighting of fixed rollouts. Recorded rewards and judgements stay unchanged. Modeled prices and weights await calibration.
Loading the saved example…
Source text
9 core · 6 disclosure · 3 repeat
Synthetic-fixture checks, not diagnostic accuracy.
case2RL synthetic fixture · Saved audit Example data (JSON)
ask_patientPatient history
examinePhysical findings
order_testsNamed investigations
diagnoseFinal answer
Responses · Provenance · Source pages · Cost
Inspect every turn →This browser · Portable JSONA published case report, two recorded rollouts, and a separate synthetic example. The PDF converter drafts case settings and checks numbers and page references. Clinical review is pending. No models run on this page.