---

name: coherence_alpha_selection description: Select a coherence boosting alpha from cached validation likelihoods without model finetuning.

Coherence Alpha Selection

Use this skill when validation examples contain full-context and premise-free candidate log likelihoods and labels. It implements the paper's one-pass validation-grid choice of alpha.

Inputs

• Validation records with full_logprobs, short_logprobs, and label.
• A numeric alpha grid.

Outputs

• Best alpha, accuracy curve, and deterministic tie-breaking notes.

Workflow

1. For each alpha, call the boosted scoring rule.
2. Compute validation accuracy.
3. Choose the highest accuracy; ties prefer the alpha with smaller absolute value, then smaller numeric value for reproducibility.
4. Return the full curve for audit.

Validation

Run tests/test_alpha_selection.py.

Limitations

The skill assumes likelihoods have already been computed by a frozen LM or a declared proxy scorer.