RGTM v0.1.0

Headline result: weight-tied message passing is free when one round suffices. RGTM solves 1-hop relation detection at 1.000 accuracy (5 seeds, zero variance) with a single shared message-automata bank, so message-passing parameters stay constant as rounds increase. RGTM(R=1) is bit-for-bit identical to a single-round cair GraphTM, which serves as the correctness oracle.

Limitations, stated plainly:

• On genuine multi-hop tasks the tied bank underperforms untied per-round banks at matched clauses (0.72 vs 1.00 on the 2-hop within_distance test). One shared bank cannot specialise per hop; this is a capacity/parameter trade-off, not an efficiency win.
• No deep or recurrent TM variant tried here solves long-range relational propagation; reported as a negative result.
• Real-data runs (MUTAG, NCI1, an LRGB probe) sit at the majority baseline; on NCI1 the cair baseline itself cannot fit the training set under a 17-config search, so no real-data claim is made either way.

Also included: controlled reproductions of two recurrent-TM failure modes (node-symbol re-encoding poison and per-round deep-supervision corruption), all experiment scripts with per-run JSONL results, and the working paper draft (PAPER.md). Requires a CUDA GPU and the cair GraphTsetlinMachine package.