---

name: reprogramming_evaluation description: Validate mechanism evidence for adversarial reprogramming proxy or full recovery experiments.

Mechanism-Faithful Reprogramming Evaluation

Use this skill after a recovery run to check whether it exercised the paper mechanism: frozen target model, universal input program, output remapping, optimizer update, numeric metric, and clean source boundary. Do not accept a high metric if these mechanism checks are missing.

Inputs

• Recovery result dictionary.
• Training trace dictionary.
• Optional source manifest paths.

Outputs

• ok boolean.
• List of errors.
• Metric and mechanism summary.

Workflow

1. Verify at least one numeric metric.
2. Verify required mechanism booleans.
3. For reduced/proxy recovery, verify parameter change and loss fields.
4. Verify forbidden original repository paths are absent from the source manifest.

Validation

Run python tests/test_reprogramming_evaluation.py.

Limitations

This helper evaluates artifact consistency. It does not independently re-run the recovery experiment.