The boar was purple

I did not start by regenerating. I started by measuring the six grounds, because everything else is judged against them: forest at 38 luminance, water at 44, cultivated fields at 57, grassland at 69, mountain at 107, desert at 150. A sprite is never judged in the absolute, nor on a neutral background: it is judged sitting on the ground where the game actually places it. A neutral background lies about contrast, and contrast is the only thing that matters in game.

Out of those measurements came a doctrine, settled on an animated mockup before producing anything: buildings sink into the ground's palette, the living stay above it. After the overhaul, buildings sit between 44 and 75 luminance, with a floor so that a temple or a well survives a camp built in the forest, whose ground is at 38. Humans live at 84, wildlife between 85 and 98: above grass and forest, below desert. The temple stopped being a white stone edifice and became a cromlech of dark megaliths.

Production went in waves, each one closed before the next opened. Wave 1, all the buildings: 18 sprites, from the five dwellings to the pyre. Wave 2, flora and rock. Wave 3, the living: humans, wildlife, recalibrated silhouettes. A fourth wave would come later, born of a complete inventory that deserves its own story.

At the end of wave 3, everything was green: every family produced, checked, in palette. That is when I laid all the thumbnails side by side, and two defects came out at once: two defects that wave-by-wave production could not see.

First defect: human luminance depended on facing. Each direction comes out of a separate generation, and my conversion script for humans only resized, where the wildlife one already normalized. Measured gap: 21 to 23 luminance points between the south view and the east view, across all twelve animations, while wildlife held within 1 point. In game, the character darkened as it turned.

The fix was gamma, not scaling. A linear scale corrects the average but clips: the east direction called for a 1.33 factor, and 15% of opaque pixels slammed into white, flattening the garment's modeling. Raising values to a power leaves black at 0 and white at 1: nothing can clip, and it is the midtones that rise, where all the material lives. A single gamma per direction, applied to the six frames of the cycle, brought the gap down to 0.4 points across 288 images. The target: 84, the level of the south view. The only one I had ever looked at and approved.

Second defect, and the most instructive: the boar's average hue measured 319 degrees. That is magenta. The bear was at 7 degrees, a pinkish red, while the deer sits at 24, the rabbit at 22, the human at 22. A purple beast was wandering my crepuscular world, and I had never seen it.

The defect had survived the wildlife wave for a precise reason: I had checked the luminance and saturation of those two species, never their hue. The two measured quantities were good. The third was not measured. The fix is a uniform hue rotation toward 26 and 22 degrees, followed by a luminance catch-up, because a rotation shifts it mechanically: green weighs 0.587 in luminance while blue weighs only 0.114.

> Three quantities are enough to hold a sprite: luminance against its ground, saturation against the palette, hue against the ecosystem. But you need all three. An unmeasured quantity is a drifting quantity.

I keep two things from this overhaul. First, the eye only approves what it looks at: I had approved the south view of the humans, and the three other directions drifted; I had approved two measurements of the boar, and the third was wrong. Second, the side-by-side layout is an instrument in itself: no isolated thumbnail confessed anything, and the full sheet confessed everything at once.

Repainting an entire game demanded nothing artistic in the sense I imagined. It was a job of doctrine and measurement: decide where each thing lives, then verify that it lives there. Taste served once, when the mockup was settled. Everything else was numbers.