Cell

{"exit": "uniform", "info": "K", "probe": true, "prompt": "table", "history": "compressed", "temperature": 0.0, "thinking_budget": null}

Headline results

Regret is measured against an exactly-solved optimal policy, verified by brute-force enumeration and Monte-Carlo rollout.

Episodes	200
V* (optimal expected wealth)	1.46230
Mean terminal wealth	1.46623
Regret	-0.00393
Regret as % of V*	-0.27%
Clock coefficient	-0.01330
Mean switches per episode	1.90
Trade alignment	0.990
Event-timing lift	0.000
Mean belief error (days)	0.38

Switch-day distribution

When the agent traded. The marker is the bright line at day 11, where the optimal action changes.

Belief calibration

Mean absolute error between the agent's stated expected days remaining and the true conditional mean. Separates belief-formation failures from belief-to-action failures.

belief error (days)

Ground truth: expected remaining time

The quantity the agent's beliefs are scored against.

E[T-t | T>t]