Distilling gpt-oss-120b's argument extraction into small Qwen adapters. v4 retains 85% of the teacher F1 on Politifact-style claims.