The world's light is continuous; a file holds only a handful of steps. Crush a smooth sky into a few brightness levels and it cracks into bands (posterisation) — unless you scatter the error as noise, and the eye reassembles the smooth tone that isn't there (dithering). Slide the levels down, toggle the dither, and watch tone get reconstructed from almost nothing.
Real quantisation run live: each channel is snapped to N evenly-spaced levels, q = round(v/255·(N−1))/(N−1)·255. With dither OFF you see the exact banding; with ordered (Bayer 4×4) dither ON, a threshold is added before snapping so the quantisation error is spread as a pattern the eye averages back to smooth. The panel measures it: the palette size (N³ colours), the mean absolute error vs the original, and how dithering keeps that error while destroying the bands.
The banana scene is the stand-in; ordered dither is one of several methods (Floyd–Steinberg diffuses error differently). The level maths, the palette count and the measured error are exact.
'Reconstructing' smooth tone from few levels is a trick played on a viewer's eye — the machine is not restoring what it saw, it is arranging error. The dithered image is my input dressed up; there is no perception behind it, no sentience.