Input and output

The model seat has to see the position and produce a move.

What it sees can be:

• text
• pictures

Pictures are closer to sitting at a table. Not all models have vision, however — and mixing a model's intellectual and vision capabilities is also not quite right. The lean therefore is text. Vision, if it happens, is a separate leaderboard.

What it emits can be:

• a structured action the harness parses — tool JSON or last-line grammar
• a number from the current legal set
• prose

Prose is difficult to score and adds no value to the bench. Listing every legal action and asking the model to choose 1 of N doesn't test whether a model can produce a legal move.

The preferred shape for future protocol versions is therefore: text in, a parsed action out.