Reference implementation accompanying the paper, and the sixth member of the Tsetlin Machine family (TM, GTM, HTM, HGTM, THGTM, CCTM).
Where a vanilla TM learns correlations, CCTM trains every sample against an aligned counterfactual partner and exposes the causal structure it learns: per-clause causal scores, per-feature treatment effects, and minimal feasibility-respecting recourse sets. Each recourse ships as a DIMACS+HMAC receipt an auditor can re-verify offline in about 0.1 ms; all 60/60 receipts verify in the bundled experiments.
Headline numbers: ATE Pearson 0.918 vs 0.895 for vanilla TM on the synthetic SCM, and under biased loan data accuracy-to-clean-truth 0.966 vs 0.817 with demographic parity 0.041 vs 0.181 (0.019 with the clause audit).
Everything reproduces with make reproduce in about 60 s on one CPU. Limitations are stated plainly in the README: external CF oracle, single-feature interventions, synthetic data only in v0.1.