AI model
—
The floor vs the classifier.
A cube is the easy case: every pick can be checked, so a model that guesses badly gets corrected for free. The interesting question is what happens in the ordinary case where you can’t check — and whether a confidence score is the same kind of thing as a check. It is not, and the cube is a clean place to show why.
symbolic floor
four subgroups · exact tables · 0 model calls
—
scramble:
Every challenger, same cube, recorded once
Each model was given this exact cube, shown the six faces, told the notation, and allowed six rounds with the updated state after each one. Every move it proposed was applied by the same kernel the floor uses, so “solved” means solved. A model we could not reach is marked as such — nothing is guessed on its behalf.
| solver | verdict | moves | thinking |
|---|
Over 24 scrambles, both solvers verified on every cube
| layer method | symbolic floor |
|---|
The part that isn’t about move count
Feed the solver a solved cube with one edge flipped in place. Every sticker is a legal colour and every piece sits in a legal slot — and the cube cannot exist. A classifier over a fixed answer set has no output for “this input is not in the world”; the best it can do is be unsure between cases that are all wrong.
Control: two flipped edges is reachable and returns REACHABLE — a check that refuses everything would be worthless.
Prompted by a public post demonstrating a classifier picking cube cases inside a fixed procedure — source. Every number on this page was measured by us, on our own implementations of both methods.