Correction Is Not Repair: Epistemic Residue and Capability-Sensitive Contradiction Repair in LLMs

This record contains a short technical working note on contradiction repair in large language model outputs. The note examines a small pilot study of “dirty” contradiction cases: cases in which a model first produces an unsupported claim, later appears to correct it, but the later correction may still contain unsupported epistemic residue. The central claim is narrow: correction is not identical to repair. Robust contradiction repair may require not only detecting a conflict and changing the answer, but also auditing the later correction, removing unsupported residue, and preserving residual uncertainty. The pilot compares responses from three models: Qwen3 1.7B, GPT-4.1 mini, and Claude Sonnet 4.5. The result suggests that residue removal is capability-sensitive. Qwen3 1.7B often corrected the main error while retaining unsupported residue; GPT-4.1 mini improved substantially; Claude Sonnet 4.5 performed best in the tested cases. This note should be interpreted as a small technical working note, not as a full benchmark or universal claim about all LLMs. Its purpose is to define the problem of epistemic residue in contradiction repair and motivate larger follow-up evaluation.