---

name: proxy_recovery_harness description: Run a bounded mechanism-faithful RAIL-KD proxy experiment with executable recovery and validation evidence.

Proxy Recovery Harness

Use this skill when full RAIL-KD GLUE transformer distillation is blocked but soft-mode recovery permits a reduced, mechanism-faithful experiment. It must call or import the generated mapping, representation-loss, and combined-objective skills rather than duplicating their contracts silently. Do not use this skill to claim full paper reproduction unless real model, data, and training evidence exists.

Inputs

• Attempt directory containing module_plan.json and environment/runtime_handoff.json.
• Generated skills root containing random_layer_mapping, rail_representation_loss, and combined_kd_objective.
• Synthetic experiment configuration: teacher layers, student layers, tokens, dimensions, epochs, seed, learning rate, and RAIL variant.

Outputs

• recovery/recovery_result.json with numeric loss_reduction_fraction.
• recovery/logs/training_trace.json with params_before and params_after.
• recovery/logs/generated_skill_invocations.json proving generated skill usage.
• recovery/logs/generated_data_item.json describing the synthetic hidden-state example.
• Mechanism checks for random mapping, coverage, pooling, normalization, objective composition, and optimizer update.

Workflow

1. Read the module plan target and runtime handoff.
2. Record that full GLUE distillation is blocked unless the handoff proves otherwise.
3. Build deterministic synthetic teacher/student hidden states.
4. For each epoch, call the mapping skill, compute the RAIL representation loss, combine CE/KD/RAIL losses, and update trainable scalar student parameters.
5. Save a command-produced result with before/after metrics and source-boundary evidence.
6. Run the Distiller recovery experiment validator after the harness completes.

Validation

Run a smoke experiment:

python scripts/run_proxy_recovery.py --attempt-dir /path/to/attempt --skills-root /path/to/generated/skills --epochs 6 --seed 11
python -m pytest tests

Limitations

The default experiment is a declared reduced proxy. It validates the RAIL-KD mechanism but does not reproduce GLUE scores, large transformer optimization, or out-of-domain results.