Sudoku: rules repair perception.
A four-clue 4×4 puzzle. The neural part reads the clues with noise; the symbolic part enforces the rules. Switch modes, turn up the noise, and watch what each part contributes.
The Sudoku lab
Board legend. Shaded cells with an orange dot are clues. The thin bar under a clue is the neural module’s confidence in its own read; red means the read disagrees with the true clue. Teal digits are filled by the symbolic solver.
Mode
What the neural part reads
| Cell | Read | p(read) | Status |
|---|
What happened
Why this is NeSy
The neural part supplies digit probabilities; the symbolic part supplies the row, column, and box constraints. The rules are not a post-hoc filter: they change which solution is selected. At low noise, the perception is reliable and the rules are almost redundant. At high noise, the rules can correct a misread — or choose a different valid solution, because four clues are not enough to identify which read is wrong.
The puzzle has four clues, the minimum for this solution. There are 288 valid 4×4 grids; the solver ranks them by how well they explain the perceived clues, then picks the best. That ranking is the interface.
Three regimes, one lesson.
Perception is enough
The neural read is correct; the solver fills the grid. The symbolic part is doing real work, but the interface is quiet.
The rules repair a read
A clue is misread, but the other clues and the rules make the true digit the best explanation. This is the NeSy sweet spot.
The rules choose a different world
With four clues, a misread can make another valid grid the best explanation. The output is still a valid Sudoku — just not the one you started with. That is not a bug; it is the limit of the information.
It is a 4×4 toy with a simulated perception module, not a benchmark. The point is to make the interface visible: probabilities in, constraints over them, a solution out. The papers in §6 use the same idea at a much larger scale.
Back to the seven worked examples · Read the formulation again