---

name: sac_off_policy_replay description: Build and sample validated off-policy replay batches for Soft Actor-Critic recovery harnesses.

SAC Off-Policy Replay

Use this skill when a SAC experiment needs a compact replay-buffer abstraction that stores transitions and produces deterministic minibatches. Do not compute SAC losses here.

Inputs

• Transition records containing state, action, reward, next_state, done, and log_prob.
• Optional integer sample indices.

Outputs

• A validated replay batch.
• Summary counts and deterministic sampled transitions.

Workflow

1. Validate every transition has the required fields.
2. Convert numeric fields to finite floats and done to bool.
3. Preserve the stored transition order.
4. Sample by explicit indices so tests and recovery runs are reproducible.

Validation

Run python tests/test_replay.py or validate the full skill tree with the Distiller validator.

Limitations

This is a bounded recovery utility, not a high-throughput production replay buffer.