onepass-c4 — a 706 k one-pass specialist that plays Connect Four

A "one-pass specialist" scores a supplied list of options in one forward pass: context bytes in, one score per option out. This one plays Connect Four. It is the second member of the family after precisit/one-pass-sv-forms.

• Shape: 224 context bytes, 8 option slots x 24 bytes (the model is shape-agnostic; this is only the export shape).
• Weights: model.safetensors (2.83 MB, 706 k parameters).
• Browser: onepass-c4-8x24.onnx — 3.9 ms per decision on WebGPU, 7.8 ms on wasm, measured in Chrome on an M4.
• Play it: https://precisit.github.io/onepass-web/demo/c4/

How well does it play

Measured, and not flattering:

metric	value
optimal-move rate (2 000 held-out positions)	78.05 %
centre-column prior (one line of code)	62.15 %
vs a depth-1 hand-written bot, model as first player	40–0
vs depth-2 / depth-4 bots	0–40

It reads the position from the move history line in its context, not from the board layout: when the board is rewritten (every stone's owner swapped, same column heights) the optimal rate moves 0.15 points, while shuffling the move history costs 16.2 points. With the history removed entirely it falls below the centre prior (54.75 %).

Provenance

Corpus and training from the c4-onepass spike in precisit/mira-spikes; toolkit in precisit/one-pass-specialists. Labels are exact solver values (the game is solved), positions with 26–38 stones.