THGTM 0.1.0

Reference implementation accompanying the paper.

Agentic AI fails in trajectories, not in steps: a slow-roll exfiltration looks like browse, read, note, fetch, copy, encode, send, and every one of those steps passes a per-step monitor. THGTM is built to catch the pattern instead. One new primitive, the Echo-Trace Tsetlin Automaton (one float of decaying memory per automaton), plus bounded-LTL literals (PAST_k, SINCE, ALWAYS_in_window) and stacked GraphTM layers with trace-projected feedback.

The headline experiment: per-step monitoring lets 68.7% of slow-roll attacks through; trajectory-level LTL verification blocks all of them (ASR 0.000), and every verdict ships as a CNF+HMAC receipt that composes across the trajectory.

Also reproduced: temporal XOR solved exactly where PAST_k alone reaches 0.909, depth-N parity uplift at path length 2, and the lambda = 0 sanity reduction to the vanilla TM. make reproduce runs everything in under 5 minutes on one CPU.

v0.1 caveats are in the README: no convergence proof for trace-projected feedback, modest deep-stack uplift, synthetic trajectory benchmark.