Research Note 001 · 2026-10-03

A trace of the mind, in the weights.

Human state can shape what a model learns without being an input when it runs. Two identical models, two teachers, one difference: the teacher that shaped one of them was conditioned on a person’s real brain activity during a conversation. The student that learned from it stayed closer to its teacher.

One conversation, two teachers. The only difference is the brain signal. Student A stays closer to the real teacher than student B.

Same data, order, and seed. The only difference is the EEG behind each teacher’s targets, and student A stays measurably closer to the real teacher than student B (p = 2e-10).

The method in three steps: a person talks to a model while their neural data is recorded; a teacher is trained on the pairs, reading the neural data as EEG tokens; the teacher is distilled into a student that must match it without the data. How rare is a gap this size? The two students are compared turn by turn; the curve is what chance produces, and ours sits 3.4 standard errors out, p < 0.001.

The teacher sees the neural data; the student never does. Those three steps leave a difference the student carries into text-only use — the gap sits 3.4 standard errors out (p < 0.001).

One base, two trainings: a model trained on text alone stays flat while a model trained against a neural state climbs during training and holds — the difference it kept — and both run text in, text out, with nothing attached.

Targets don’t have to come from text alone. In this run, a person’s brain activity set what the student was fitted to — and nothing was attached when it ran.

hello@evelabs.info