RLT is getting attention, so I tested a small implementation (79K params, 3 seeds). Two synthetic tasks, trained at 32 steps. Final-state accuracy: 32: 100% / 100% 128: 60.8% / 20.7% GRU: ~100% at…