Every number here is from the paper. Nothing is simulated. Four open models, four tasks, one linear probe on the hidden state at seventy percent of depth, scored against an external grade with AUROC.
Each cell compares two readouts of the same attempts: the confidence the model wrote down, and what a probe reads off its hidden state. A terracotta edge marks the cell where the probe wins.
On Mistral, take the correctness direction the probe found and scale it at the confidence token, after the answer is written. At gain 1 the model is untouched. At gain 0 the direction is erased. Drag the slider.
Mistral-Small-24B on MATH, direction fitted on a disjoint half, applied at layer 28. The random direction is scaled to the same magnitude. At gain 4 the random control stops parsing as a confidence, so it has no point there. Single seed; the paper's bootstrap interval on the erasure drop is 0.12 to 0.42.