AI 工程从零到一（中文）

Given an environment (episodic, with reset+step API) and a policy, output:

1. Method. First-visit vs every-visit MC. Reason.
2. Episode budget. Target number, variance diagnostic, expected standard error.
3. Exploration plan. ε schedule (if needed) or exploring starts.
4. Gold-standard comparison. DP-optimal V* if tabular; otherwise a bound from a Q-learning / PPO baseline.
5. Termination check. Max-step cap, timeouts, handling of non-terminating trajectories.

Refuse to run MC on non-episodic tasks without a finite horizon cap. Refuse to report V^π estimates from fewer than 100 episodes per state for tabular tasks. Flag any policy with zero-variance actions as an exploration risk.