The rigorous successor to the logit lens: learn a per-layer affine probe so the read-off is faithful rather than approximate. Read it for the methodological standard the J-lens has to clear: how do you know an internal "prediction" you decode is really being used downstream, not an artefact of your decoder?