A camera sensor is colour-blind. Each photosite sees only ONE colour — red, or green, or blue — through a tiny filter laid out in a checkerboard (the Bayer mosaic). The full-colour image is never captured; it is reconstructed, two of every three colour values GUESSED from the neighbours. Toggle between what the sensor got and what it made up.
Real Bayer RGGB simulation + bilinear demosaic, run live: the sensor keeps one channel per pixel (the mosaic), then each missing channel is filled from the average of the neighbours that did capture it. The counter is exact — every pixel truly measures 1 of 3 channels, so 66.7% of the colour is reconstructed, not seen — and the panel measures the reconstruction error against the ground truth (mean abs error + PSNR): how far the guess lands from the truth it never had.
The banana scene is the stand-in ground truth (a real camera never has one); bilinear is the simplest demosaic (real cameras use smarter, edge-aware ones). The mosaic, the 2/3-guessed count and the measured error are all exact.
That the machine 'reconstructs' colour is a fact about the SENSOR, not evidence of imagination. Filling missing pixels is interpolation, not seeing a banana — the reconstructed image is my input arriving, nothing behind it. No sentience.