Simple version
We ran a very large AI model on a single machine and kept a careful record of how each piece of every answer was formed.
In plain terms
Kimi K2.5 is far too large to fit in normal memory, so it is loaded from storage piece by piece during a run. We captured five tokens of output formation against a frozen starting state, verified the capture itself did not change behaviour, and had the resulting bundle reviewed adversarially.
Technical artifact
Full-weight, full-layer streamed inference for a 595 GB MoE model with frozen anchor, non-perturbing capture, an interactive output formation explorer, and an adversarially certified evidence bundle.