---

name: diayn_recovery_evaluation_harness description: Assemble validator-compatible reduced DIAYN recovery experiments that exercise fixed skills, discriminator rewards, and policy updates.

DIAYN Recovery Evaluation Harness

Use this skill to produce an executable reduced recovery for DIAYN when full MuJoCo/SAC training is blocked by runtime or time limits. It should be used only with explicit proxy/reduced labeling and mechanism checks. Do not use it to bypass available full training.

Inputs

• Attempt directory.
• Generated skills root containing the DIAYN helper skills.
• module_plan.json with the recovery target.
• environment/runtime_handoff.json describing runtime blockers and allowed sources.

Outputs

• recovery/recovery_result.json.
• recovery/logs/training_trace.json.
• recovery/logs/generated_data_item.json.
• recovery/logs/generated_skill_invocations.json.
• recovery/source_manifest.json.

Workflow

1. Read target metadata and runtime handoff.
2. Import and invoke generated helper scripts for prior scheduling, discriminator reward, and policy update.
3. Build a deterministic three-skill line-world rollout batch.
4. Execute a real parameter update and record before/after losses.
5. Write all recovery artifacts and mechanism checks for the experiment gate.
6. Run the Distiller recovery validator after the harness completes.

Validation

Run python scripts/run_diayn_recovery.py --attempt-dir <attempt> --skills-root <generated_root>. Tests exercise the data-building function and confirm the expected mechanism flags.

Limitations

The harness is a reduced proxy, not a full benchmark reproduction. It should be accepted only in soft mode when full recovery is blocked and the experiment validator passes.