ConfTM 0.1.0

Reference implementation accompanying the paper. Ninth member of the Tsetlin Machine family.

A TM gives you a label; ConfTM gives you a set of labels with a finite-sample promise attached: the true label is inside with probability at least 1-alpha, no asymptotics, no distributional assumptions beyond exchangeability. The wrapper is thin and model-agnostic across the TM family, with three TM-native nonconformity scores.

Three calibration modes, all validated in the bundled experiments:

• Split conformal: marginal coverage holds at every tested alpha.
• Mondrian: per-class coverage [0.95, 0.90, 0.97, 0.90] at alpha = 0.10.
• Adaptive (ACI): 0.900 +/- 0.0003 coverage under covariate drift where static calibration decays to 0.836.

Plus coverage-guaranteed counterfactual recourse: flip sets that keep working under feature noise (91.3% robust flip rate at alpha = 0.10).

make reproduce runs all four experiments in about 3 minutes on one CPU. Limitations are stated in the README: label-set vs argmax semantics, exchangeability, synthetic data in v0.1.