Each work is a pair. The left panel is a photograph as it was made. The right panel is the same photograph carrying one intervention derived from its own measurements. Side by side, the difference often lives at the threshold of vision, and it is stated exactly, down to fractions of a single tonal step. The numbers are the work, and anyone may measure them again. You are offered almost nothing to see, and everything to check. · · · In the fraction of the frame where the file is clipped, the tone now runs backwards from each region's own edge inward, through the places where the file stopped describing the scene. This file is clipped at both ends. 5,358 pixels sit at pure white in 577 regions, and 5,083 sit at pure black in 2,341, which comes to 2.3610% of the frame where the photograph stops describing the scene and instead asserts exactly 255 and exactly 0, the two values a file falls back on when it has run out of words. The blind regions were not filled in. They were marked, by being made to behave impossibly. Every clipped pixel moves toward its own nearest surviving neighbor, minus five levels in the highlights and plus five in the shadows, weighted by its distance from any pixel that still knows something. The depth of the reversal is the depth of the blindness. All eleven blown regions of nine pixels or more now darken as you move inward from their own edge. Every one of them brightened before. A simple specular highlight should not be dimmest at its center. The rule surfaced an asymmetry that belongs to the file rather than to the treatment. The shadow clipping is only 2.2 pixels deep at its worst, scattered dark speckle rather than genuine blindness, so under the same rule it earns an average lift of only 2.45 levels. The cost is eight levels of 255 at the worst pixel, and 0.0906 of a single level averaged across the frame. The file is grayscale to zero tolerance, and the change was applied identically to all three channels to keep it that way. Presentation. The mount is neutral 254, one level below white, and the exact threshold that divides the panels. The untouched photograph holds 5,358 pixels above it. The altered one holds none. Anything in the print lighter than its surround is a value the file clipped at its ceiling, so the frame is the instrument that reads the difference. The rule is 137, the neutral of the panels' own mean. To check. 14,518 pixels move, none by more than 8 of 255, and the file stays perfectly grayscale. Download this work's original file, 03_2.3610_percent.png, not a screenshot or a marketplace preview, attach it to an AI that can run code, and paste this: Run this exact line on the attached file, changing only the file path if needed. Show me its exact output, and say whether that output is exactly: peak 8 changed 14518 spread 0 python3 -c "import numpy, PIL.Image as I; c=numpy.asarray(I.open('03_2.3610_percent.png').convert('RGB')).astype(int); a=c[600:2595,600:2595][::3,::3]; b=c[600:2595,2895:4890][::3,::3]; d=abs(b-a); print('peak', d.max(), 'changed', (d.max(2)>0).sum(), 'spread', abs(b[...,0]-b[...,2]).max())"