What did symmetry hide?
A picture goes in. Rotations and reflections overlap. At every pixel, the brightest copy wins. Can we recover the picture, or only invent one that makes the same mandala?
The forward machine, opened up
Choose a scene and press Try inverse. The solver sees only the rendered mandala(s), not the original or the winning-copy map.
The browser inverse is a fixed-iteration, smoothness-regularized gradient heuristic. The native Python benchmark uses L-BFGS-B. Both are candidates, not proof of unique recovery. Browser rotations use continuous bilinear interpolation; OpenCV's interpolation table causes small differences.
Fit is not identity.
Forward agreement. The candidate almost perfectly reproduces the mandala.
Actual source recovery. The face reconstruction remains substantially wrong.
Actual source recovery improves when we change how the scene is interrogated.
What if the folds are questioned differently?
From the Varjoluotain shadow problem: change the measurements before the maximum hides weaker copies. Every primary arm observes three full-size grayscale images. The fixed masks are digital interventions on each transformed layer, not an actual cardboard plate.
Loading independently reproduced results…
Best average source recovery on this fixed pilot.
Better than one N=16 composite, not better than three symmetry views.
Maximum observed change between two originally indistinguishable source images. Ordinary N=3, 7, 16 show zero difference.
Winner-selected source-pixel coverage (coded / extra symmetries): Choose an individual scene. Coverage is not unique invertibility. Contrast inversion also makes the exact collision visible, but performs poorly on these images with the fixed solver. The predeclared experiment distinguishes stronger sensing from merely generating a plausible inverse.
Two different originals.
One identical output.
One all-white source and another with a single missing pixel yield the same mandala at N=3, 7 and 16. At N=1, they're distinguishable.
No decoder, no matter how sophisticated, can infer which source was used from those equal observations alone.
Folding ≠ revealing.
Rotation and reflection move evidence around. The max operation suppresses all nonwinning contributions. Provenance reveals which branch won, not what the other branches carried.
Inverse rendering can select a plausible source. Identifiability requires enough independent measurements to rule out alternatives.
Read the frozen protocol ↗