---

name: selection_ablation_evaluation description: Compare Moderate-DS median selection against center-close, far-from-center, and two-end score policies.

Selection Ablation Evaluation

Use this skill to test whether a selected coreset follows the paper's intended balance between centrality and diversity. It is appropriate for synthetic fixtures, representation datasets, and recovery harnesses that already have per-record scores. It should not compute hidden representations or replace full downstream model evaluation.

Inputs

• Score records with stable id, label, and numeric score.
• Selected ids for one or more policies.

Outputs

• Per-policy mean score, score spread, class coverage, balance score, and advantage diagnostics.

Workflow

1. Build baseline selections for center-close, far-from-center, and two-end policies when requested.
2. Compute centrality as closeness to the full score median and diversity as selected-score spread.
3. Combine these into a deterministic balance metric suitable for fast proxy recovery.
4. Report whether the moderate policy improves over the strongest extreme baseline.

Validation

Run python scripts/evaluate_selection_policies.py --self-test. The smoke fixture is designed so the median policy has a positive mechanism-faithful advantage.

Limitations

The balance metric is a proxy diagnostic, not a replacement for CIFAR or ImageNet test accuracy.