Reproduction of "Evaluating Bivariate Causal Statements Based on Mutual
Compatibility" (OpenReview b3EvCd8sYE).
Theory claims are settled exactly: Lemma 2.3 symbolically, the Definition 2.7 path-tracing decomposition to machine precision, the NP-hardness reduction by exhaustive enumeration of all DAGs at n=3,4, and Lemma 3.7 over the complete 512-graph universe at n=3. Each is paired with a destructive control.
The LLM half of claim 4 is a clean-room test on ten open models (864 executed calls), since the authors' Bedrock pipeline and data are absent from the release.
python3 claim1.py && python3 claim2.py && python3 claim3.py
python3 llm_scores.py