Detecting LLM Agent Contradictions Using NLI and Total Variance — A Python Implementation

LLM agents can produce logically contradictory outputs across repeated runs of the same query, a failure mode worse than simple variance. The author built a middleware layer using the Total Variance formula from arXiv:2602.23271 and NLI contradiction detection to identify when outputs are not just different but logically opposite. The implementation uses sentence embeddings with L2 normalization to compute pairwise variance scores across answers, findings, and citations, deriving a reliability score.