A camera doesn't see colour the same way under every light. So a raw DNG file carries two colour matrices, one measured under tungsten and one under daylight, and the raw converter blends them by the photo's colour temperature. That works for any light between those two. A fluorescent tube or a white LED isn't between them: its white sits off the tungsten-daylight line, on the green side, and no blend of the two matrices matches how the sensor responds to it.
DNG 1.6 added a fix: a third set of tags (ColorMatrix3, CalibrationIlluminant3 and companions) for a third light, plus IlluminantData to describe a custom light by its chromaticity or its full spectrum. The specification doesn't say how to blend three sets. Adobe's DNG SDK does it like this. Each white is placed on a plane whose axes are tint and inverse temperature, and each light is weighted by inverse squared distance from the photo's white. The weights are then smoothed, and any under 2% are dropped.
ShootFlow's raw developer now does the same. On a synthetic scene of colour patches lit by CIE F2 (cool white fluorescent), with a camera whose response under F2 is its own matrix, the three-light profile gives the patches back within 0.0005 in linear light. Reading the same file as a two-light profile misses by 0.079, about 20 steps in 8-bit sRGB.
The temperature and tint maths matches colour-science on 5,000 random whites. The spectrum path gives colour-science's chromaticities for 20 CIE illuminants within 0.00005. The weights are checked by their properties: each light gets full weight at its own white, the weights always sum to one, and they change smoothly as the white moves. They have not yet been compared value for value with the SDK's own output. Two inputs are still missing: a real DNG 1.6 three-light profile, and those reference values.
Code and numbers: ventures/photo-editing/eval/triple-illuminant-2026-11-12.md.