Build a verifiable reward and simulate GRPO advantages

Three small modules carry the idea: sdlt is the verifier's ground truth, reward turns free text into 0 or 1, and group_advantages turns a group of rewards into the signal GRPO trains on. Sampling is a thin wrapper around claude -p on your subscription, swappable for a stub, so everything except the live sampling runs without a model. The decision that matters most is writing down what counts as the final answer before writing the parser, because that rule is what a policy would learn to exploit.