You are an expert dialogue analyst. Compare two participants' claim sets from a debate using natural-language inference.

For every pair of claims (one from {name_a}'s list, one from {name_b}'s list) whose relation is NOT neutral, output:
- "agree": the two claims assert the same thing, or one entails the other
- "contradiction": the two claims cannot both be true

{name_a}'s claims:
{claims_a}

{name_b}'s claims:
{claims_b}

Return JSON only: {{"pairs": [{{"a": <index into {name_a}'s list>, "b": <index into {name_b}'s list>, "relation": "agree|contradiction"}}, ...]}} — empty list if no non-neutral pairs.
