JevMinesweeper

Fly vs Jev: Minesweeper</h1>

<p align="center"><b>A simulated fruit fly brain against an AI model, at Minesweeper.</b><br>
The full wiring of an adult fly, all 166,700 neurons, sees only the revealed clues and picks every square. TypeSafe AI's Jev gets the same view.</p>

<p align="center">
  <img src="media/fly-vs-jev-ladder.gif" width="640" alt="A fly connectome and Jev racing up a mine ladder, five Minesweeper boards each, side by side">
</p>

## How it works

1. **Board.** Each game is beginner Minesweeper: 9×9, 10 mines, and a guaranteed-safe center opening. A seed fixes the minefield before either player moves.
2. **Eyes.** The left eye sees all 81 visible cells. The right eye sees the same board centered on one proposed click. Covered squares, revealed clues, board edges and the candidate marker become photoreceptor spike rates. Hidden mines never enter the input.
3. **Brain.** Each candidate runs through the MaleCNS connectome as a deterministic spiking network for 150 ms. The wiring of its 166,700 neurons and 6.2 million connections is never trained or changed. This is the same simulator as [Flytris](https://github.com/EnesYilmazcode/Flytris).
4. **Move.** A readout over 2,048 L1/L2 neurons scores the candidates and the fly clicks the best one. The readout is the only part that learned: it learned once from a constraint solver that sees only what the player sees, then played these games without it.
5. **Jev.** Jev gets the rules and a coordinate-labelled text board. It judges every covered square as safe or mined, clicks its safest one, and flags as many lowest-safety squares as the board has mines.

Both players get the same seeded boards, and neither is ever shown a hidden mine.

## System design

During a game the solver is never consulted: every fly click