RobTM 0.1.0

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

Most robustness work reports an accuracy number under attack and stops. RobTM ships the other thing you actually want: a per-sample certificate. For every input it computes the exact minimum number of feature flips that would change the prediction (about 0.05 ms per sample at d=10), and a CertifiedPredictor that simply refuses to answer below your chosen radius.

This release is deliberately honest about what works and what does not:

• The certificate and abstention machinery work: on XOR-pairs the selective predictor answers 34% of queries at 83% selective accuracy at tau=2.
• Adversarial training does NOT rescue dense tabular tasks: 276 of 277 medical-risk samples sit at radius 1, and both standard and robust TMs collapse under a single flip. Section 6 of the paper proposes three sharper mechanisms for v0.2.

Also included: RobConfTM, composing certificates with ConfTM prediction sets. make reproduce runs the three experiments in about 5 minutes on one CPU.