Independent AI research. Evidence you can inspect.

Run as a feedback loop, a small model’s numerical behavior carried messages in both directions, held a written bit across a fold, and ran stateful AND and OR gates on it.

Temperature and steering move a soft-token loop between repeating text and chaos. Where the loop is chaotic, rounding from how requests are batched can change what it writes.

Layer similarity reveals structure in the fitted lens; it does not establish a shared causal mechanism.

The comparison spans model families, so size alone cannot explain the differences.

Delivered edits scale with dose, but the model’s downstream response departs from that scaling.

A matched clean reference exposes a signed steering footprint; an isolated post-steering state does not. The follow-up that compared related features failed its replay gate, so those comparisons are exploratory.