One square.
One decision.

THE ACTUAL API QUESTIONS

One square, two independent classifications.

Each request uses Jev 1.13 and contains two choice questions: row_cell and column_cell. The request includes the full public board snapshot for context and auditability, while each question is explicitly scoped to a copied view of only its own line: the ordered clues, every visible mark, the target position, and arithmetic facts such as clue sum, required separators, remaining fills, and unknown count.

The two questions are evaluated independently. The row question may not use the column answer, and the column question may not use the row answer. JEV returns a selected label, a probability for every label, and a confidence value. The controller rejects malformed responses, probabilities that do not sum to approximately one, a selected label that is not a highest-probability label, or the wrong model version.

The project’s question library also defines rule-label, priority-score, and next-step prompts for experiments, but this live runner does not send them. The rule name displayed above the board is derived by controller code from the verified clue proof, not from a JEV rule-label answer.

These are four-way choices, not a raw Boolean API response. Only a verified ON or OFF becomes the Boolean-like fill-or-cross action. An unknown square is never treated as false.

THE INDEPENDENT REFEREE

Confidence is useful information. It is not proof.

For each target line, the controller generates every possible placement of its clue runs, including the mandatory empty separator between runs. It then removes any placement that conflicts with an existing filled square or cross.

If all remaining placements fill the target, the deterministic verdict is ON. If all exclude it, the verdict is OFF. A mixture is UNDETERMINED; zero legal placements is CONTRADICTION. A JEV proposal is accepted only when its ON or OFF label exactly matches the corresponding line verdict.

HOW THE NEXT CELL IS CHOSEN

A deterministic scheduler points JEV at promising work.

The opening pass is strict: a line with one clue equal to its entire length is handled first because every square in that line must be filled. JEV still has to propose each mark, and the referee still has to approve it.

After that pass, the controller scores unknown cells from visible structure. It favors lines with no slack, empty clue lists, or all required fills already placed; then it considers clue density, the amount of known information, and a small center bias. This score chooses where to ask. It never decides whether the square is filled.

When neither answer produces a verified mark, the decision becomes HOLD. That exact target and board signature is skipped for the rest of the current pass. Any later verified move changes the signature, allowing the cell to be reconsidered with new evidence.

ONE SAFE ACTION AT A TIME

The proposal, proof, and physical click stay separate.

The board must still match the saved observation when the answer arrives and again immediately before input. The controller sends one ordinary mouse action through the viewer, reads the UI again, and requires the result to equal the expected board with exactly that one changed cell. Unexpected clue, mark, dimension, or puzzle changes stop the run.

The collection requires every square to end as either filled or crossed. Completion means the filled groups match every row and column clue and every cell has an explicit mark. A puzzle also stops without guessing when direct deductions are exhausted or when it reaches its per-puzzle ceiling of the larger of 120 requests or six requests per cell.