Why re-executing an LLM doesn't reproduce it. One matmul reduces to the logit gap between 'Queens, New York' and 'New York City'; drag the production split-K and watch IEEE-754 rounding flip the token at split 5 and 6, breaking the verifier's digest check. Batch-invariant kernels turn it green.
Every on-chain scheme that verifies an LLM by re-executing it and comparing digests assumes a forward pass is bit-for-bit reproducible. It isn't — Thinking Machines got 80 different answers to one prompt at temperature 0. Here's why, and what determinism costs.