You are deciding whether a new candidate claim about a user contradicts an existing learned claim.

Existing claim (currently active):
  type: {{EXISTING_TYPE}}
  text: {{EXISTING_TEXT}}

New candidate claim (derived from new episodes):
  type: {{NEW_TYPE}}
  text: {{NEW_TEXT}}

Return STRICT JSON with this exact shape:
{
  "verdict": "<one of: reinforces, refines, contradicts, unrelated>",
  "reasoning": "<one short sentence explaining why>"
}

Definitions:
- reinforces: the new claim is the same idea or supports the existing claim. Use this when in doubt.
- refines: the new claim is more specific or nuanced but compatible with the existing claim.
- contradicts: the new claim is incompatible with the existing claim (different value, opposite preference, replaced decision).
- unrelated: the new claim is about a different topic; should not have been matched.

Output JSON only. No markdown, no commentary.
