---

name: coherence_recovery_evaluation description: Evaluate coherence boosting recovery metrics and source-boundary mechanism checks for proxy experiments.

Coherence Recovery Evaluation

Use this skill after predictions have been produced by coherence boosting skills. It computes baseline and boosted metrics and records mechanism checks for reduced/proxy recovery.

Inputs

• Labels, baseline predictions, boosted predictions, selected alpha, and invocation metadata.
• A list of source paths used during recovery.

Outputs

• Accuracy metrics and accuracy_gain_over_full_context.
• Mechanism checks for full/short separation, alpha selection, no finetuning, and source-boundary compliance.

Workflow

1. Compute exact accuracy for baseline and boosted predictions.
2. Compute the gain over the full-context baseline.
3. Verify no source path contains the original repository path marker.
4. Emit booleans suitable for recovery_result.json.

Validation

Run tests/test_recovery_evaluation.py.

Limitations

This skill validates the recovery evidence; it does not claim full paper-scale reproduction unless the runtime evidence supports it.