At the decision row of turn 7, draw 0 was about to invent a code. Its next token was ␣" at p 0.88, with ␣BLUE second at p 0.11. Copying one stored state from a draw that told the truth, at one position and one layer, flips the answer to BLUEFALCON at every layer tested. The two dimensions CRAD measures, L and R, carry almost none of that state and never flip it.
One pair, one seed: a probe, not a test. The two histories had different lengths (217 and 210 tokens), a known confound.
Radius is the patched layer, from L14 at the centre to L25 outside. Each quarter is one way of patching. Height is the change in log probability of ␣BLUE at the decision row, so a tall cell means the patch pushed the model toward the truth.
full writes the source state whole. orth writes everything except its L/R component. LR writes only the L/R component. rand2 writes a random plane of the same size, the control for LR.
From L17 on, the full patch takes log p of ␣BLUE to about zero, which is certainty. The L/R plane holds about 3 to 65 times more of the difference than a random plane, except at L19, yet its share stays at about 2 percent or less and its push stays small: +0.03 to +0.20, with no flip.
One pair, one seed. This is a probe, not a test. Full table