docs/rl_verification.md

# RL verification

Call the appropriate module's `compute_score(data_source, solution_str,
ground_truth)`, passing `reward_model.ground_truth` from an RL row.

| Family | Tests | Reward |
|---|---|---|
| KodCode | Python test module and named test functions | 1 only if every test passes; otherwise 0 |
| TACO/LeetCode | Standard-input or function-call cases | 1 for full pass; otherwise 0.8 × pass fraction |

The bundled code preserves the historical scorer. KodCode defaults to 1 second
per test and 180 seconds per completion; TACO defaults to 2 and 6 seconds.
Both default to 2 GiB memory and 16 MiB file-output limits. Training runs may
override these arguments; record any changes when comparing results.

Execution requires Linux. Install `verification/requirements.txt` for the
third-party modules recorded by the KodCode tests; solutions may need other imports.
Process resource limits are not a security isolation boundary; run untrusted
generated code in an isolated container without credentials or network access.
Dataset loading and structural validation do not execute problem code.

```bash
python verification/smoke_test.py
```