Evaluation execution
Six-policy comparison
Termination and correctness are separateTrace investigation
Each timeline step shows the policy-visible state and the harness response. Gold labels and evaluator annotations are kept in the separate evaluator-only section below.
Evaluator only · not available to a policy
Gold outcome and scored result
The gold terminal action and evidence annotations below were used only after the episode. Scenario IDs are report navigation labels. Unrequested fixture observations are not shown to the policy in any timeline.
What is deterministic here?
Candidate eligibility, evidence masking, budgets, fixture responses, evidence checks, proposal acceptance, and terminal accounting are harness behavior. Learned policies choose from their scored candidates. Fixed order follows its documented tool sequence and is deterministic.
“Completed” describes episode termination only. It does not mean the outcome matched the evaluator. Supported diagnoses and correct reviews have explicit denominators in the comparison cards. Candidate scores, native probabilities, confidence, and action probability are recorded metadata, not calibrated confidence or permission to act.
GLiClass and Laya receive different native input formats, so this comparison does not isolate model architecture. These inspected synthetic regression scenarios are not a production accuracy estimate.