---

name: coherence_context_protocol description: Build and validate full-context and premise-free-context records for coherence boosting answer-selection experiments.

Coherence Context Protocol

Use this skill when preparing multiple-choice or cloze examples for coherence boosting. It separates a long premise-bearing context from the premise-free prompt that captures answer-format or answer-prior effects.

Inputs

• Examples with premise, prompt, candidates, and label.
• Optional metadata identifying dataset and split.

Outputs

• Records with full_context = premise + prompt, premise_free_context = prompt, unchanged candidates, and label.
• Validation errors for missing fields, empty candidates, invalid labels, or leaked premise text in the premise-free context.

Workflow

1. Normalize whitespace but do not rewrite candidate text.
2. Concatenate premise and prompt for the full context.
3. Keep the premise-free context free of the premise.
4. Validate candidate count and label bounds.
5. Pass records to scoring modules; this skill must not choose answers.

Validation

Run python tests/test_context_protocol.py or validate through validate_skill_tree.py --run-tests.

Limitations

This skill does not call a language model. It prepares the contexts whose likelihoods are later computed or supplied by a cache.