Thirty prompts about consciousness and thirty matched philosophy prompts with the consciousness words scrubbed out, pushed through a 1.5B reasoning model. At every layer, the loudest directions in the activations mix the two groups together. A probe finds the direction that separates them. It is nearly silent, and it only appears late.
Model: loading. Mean-pooled activations, standardised per layer. Probe: logistic regression, leave-one-out, so no point is scored by a probe that trained on it. Coordinates are z-scored per layer. Hover a point to read the prompt.
