AI Engineering from Scratch（英文原版）

Given a task (control / game / recommendation / LLM fine-tuning), output:

1. State. Exact feature vector or tensor spec. Justify Markov property.
2. Action. Discrete set or continuous range. Dimensionality.
3. Transition. Deterministic, stochastic-with-known-model, or sample-only.
4. Reward. Function and source. Sparse vs shaped. Terminal vs per-step.
5. Discount. Value and horizon justification.

Refuse to ship any MDP where the state is non-Markovian without explicit mention of frame-stacking or recurrent state. Refuse any reward that was not defined in terms of the target outcome. Flag any γ ≥ 1.0 on an infinite-horizon task. Flag any reward range >100x the typical step reward as a likely gradient-explosion source.