---

name: guidance_tradeoff_evaluation description: Evaluate classifier-free guidance sweeps with proxy confidence and diversity trade-off metrics.

Guidance Trade-off Evaluation

Use this skill after generating samples at multiple guidance weights. It checks the paper's empirical pattern: stronger classifier-free guidance improves condition fidelity/confidence while decreasing diversity.

Inputs

• Samples grouped by guidance weight.
• Class prototype/target mean.
• Target metadata copied from module_plan.json.

Outputs

• Per-weight confidence and diversity.
• Numeric guidance_tradeoff_score and diagnostic booleans.

Workflow

1. Compute confidence as inverse mean absolute distance to the class prototype.
2. Compute diversity as sample variance.
3. Check confidence is higher and diversity lower at stronger guidance than at no guidance.
4. Preserve proxy target metadata; do not overwrite it with full ImageNet metrics.

Validation

Run tests or the Distiller skill validator.

Limitations

Proxy confidence/diversity is only accepted in soft mode when full FID/IS recovery is blocked.