FULL mess3 (paper-exact a=0.85,x=0.05, seed 42, 1e6 steps) results: repro (R2 0.987, heldout-beliefs 0.977, shuffle collapses 0.67->0.02, at Bayes floor), battery (random-init 0.938 vs trained 0.987; last-4-token baseline 0.965), positive control (belief clamp provably swaps probe readout yet next-token TV moves only 0.477->0.426 — probe basis not the computational basis), fractal + emergence simplex PNGs, full checkpoints.… See the full description on the dataset page:
https://huggingface.co/datasets/latkes/belief-state-probes-results-mess3-v1.
