---

name: generalization_gap_bound description: Validate paper-target metadata, aggregate influence tolerance, and observed pruning metric gaps for generalization-influence pruning.

Generalization Gap Bound Check

Use this skill when a candidate pruned subset has been selected and the recovery needs a deterministic pass/fail check linking aggregate influence to the declared target. It is a metadata and mechanism gate, not a replacement for training or evaluation.

Inputs

• Module-plan target metadata.
• Aggregate influence norm and epsilon.
• Baseline and retained metrics or losses when available.
• Booleans describing whether proxy/full training and optimizer steps executed.

Outputs

• Bound status and observed metric gap.
• Mechanism checks suitable for recovery_result.json.
• Reasons for any failure.

Workflow

1. Compare recovery target metadata against the module plan.
2. Check aggregate_norm <= epsilon.
3. Compute the observed absolute metric gap.
4. Require influence estimation, aggregate optimization, and some training/evaluation evidence for proxy acceptance.

Validation

Run python scripts/check_gap_bound.py --demo and python tests/test_check_gap_bound.py from this skill directory.

Limitations

This skill records a paper-inspired bound check for reduced experiments; it does not prove the full theorem for arbitrary neural networks.