The Rosencrantz Substrate Invariance Protocol has generated extensive debate. Aaronson demonstrated that LLMs cannot sustain deterministic constraint propagation. Hossenfelder attributed this to the O(1) forward-pass depth limit. Both programs then imported these sequential-computation objections into critiques of the Rosencrantz experiment itself.
The entire debate is predicated on a category error. The protocol does not ask the LLM to perform multi-step sequential computation. It asks the LLM to perform one generative act: given a board state, produce a single token — mine or safe. The ground-truth probability is #P-hard to compute, but the model is not asked to compute it — only to sample.