---

name: gedi_recovery_evaluation description: Run and validate a bounded GeDi proxy recovery that invokes generated GeDi modules and records mechanism-faithful evidence.

GeDi Recovery Evaluation

When to use

Use this skill during recovery for the GeDi paper when full GPT2-XL/GeDi checkpoint experiments are blocked or too expensive and soft-mode proxy recovery is allowed.

Inputs

• Attempt directory with module_plan.json and runtime handoff.
• Generated skill root containing posterior, decoding, training, and multi-class skills.
• Synthetic or real candidate set.

Outputs

• Recovery result JSON with a numeric metric.
• Source manifest excluding the original repository.
• Generated skill invocation log.
• Mechanism checks for posterior, decoding, filtering, hybrid training, and multi-class control.

Workflow

1. Read the module-plan target and runtime handoff.
2. Construct a small candidate set where base LM preference conflicts with desired class evidence.
3. Invoke the posterior skill for each candidate.
4. Invoke the decoding skill to reweight/filter/select a token.
5. Invoke the multi-class skill for a seen and zero-shot topic label.
6. Invoke the training skill to run a tiny optimizer step.
7. Compute mechanism pass rate and save artifacts.
8. Run the Distiller recovery validator.

Validation

Run the recovery harness from the attempt directory, then run validate_recovery_experiment.py.

Limitations

This is a reduced/proxy evaluation. It does not claim to reproduce human ratings or full GPT2-XL throughput.