Swat.

Catch it. If you can.

A browser arcade game with a tiny airborne fly, an oversized swatter, and a live view of the neural circuit behind its escape reflex. Move the swatter closer and watch Tiny mind light up before you even click.

Play Swat · How it works · Neural model · Contributing

Swat runs a simplified spiking model of 6,000 traced fruit-fly neurons and 1,275,994 directed connections from the MaleCNS v1.0 connectome. The connections preserve 10,659,685 original synaptic contacts. These are different counts: several contacts can connect the same pair of neurons.

Simulated vision feeds an extracted LC16-to-MDN retreat circuit. Its activity influences evasive turns and bursts; the brain viewer displays the model's actual recent activity. Baseline cruising, flight animation and collision rules are authored game mechanics.

This is a playful adaptation of a real circuit, not a whole-brain emulation, a validated flight controller, or a trained AI. The game uses one fixed, fully enabled circuit. Smaller graphs and neuron-silencing interventions are retained for offline tests, not player difficulty controls.

Use Node.js 24 and npm. No API keys, account, database or inference server are needed.

git clone https://github.com/hrook1/Swat.git
cd Swat
npm ci
npm run dev

Open the local address printed by the development server. The prepared circuit files are included; a normal build does not download the raw connectome or require Python.

npm run verify   # Types, lint, formatting, secret checks, tests and build
npm run preview  # Serve the production build locally

Read the welcome screen, click Let's go, then start a 30-second round. Aim ahead: the swatter lands 240 ms after your click, at the position you committed to. Another swing becomes available 560 ms after the previous click.

Reset returns to the original seed; Rematch generates a new route. A ?seed=34 URL fixes the starting seed for a reproducible setup, not a replay of someone else's pointer movements.

Record clip captures the game and neural view together, silently, at up to 30 fps. Stop manually or let it finish after 30 seconds. The browser chooses supported MP4 or WebM output; a PNG snapshot is the fallback. After a round, Slow-mo replay shows the last swat—or recent gameplay if no swat was taken—at 0.4× using recorded pose and activity snapshots.

Tests sit beside the behaviour they protect. React and Three.js handle presentation; a dedicated Web Worker owns neural integration and game state. The production output is a static Vite build that can be hosted on Vercel without a backend.

The automated suite checks deterministic simulation, numerical reference cases, graph integrity, input/output interventions, worker recovery, swat timing, hitbox projection, replay presentation and UI contracts. An optional longer-running check is available:

RUN_GAME_SOAK=1 npm run test:unit -- src/game/Game.soak.test.ts

These are software and model regression tests, not biological validation or a guarantee of performance on every device. The large graph costs memory and CPU; the UI reports when simulation speed falls behind. A recent browser with WebGL and hardware acceleration is recommended. Video export support varies by browser. See the architecture notes for runtime boundaries and a release checklist.

The adapted MaleCNS v1.0 data comes from FlyEM / HHMI Janelia Research Campus, the University of Cambridge, the MRC Laboratory of Molecular Biology and Google Research. It is provided under CC BY 4.0. Selection rules, source checksums, anatomical omissions and modelling assumptions accompany each prepared graph.

• MaleCNS data and downloads
• Data attribution and adaptations
• Model details and data reproduction
• Sen et al.: visually evoked retreat
• Shiu et al.: connectome-based point-neuron modelling

The upstream researchers have not endorsed this game. Their data license is separate from any license applying to the application code or other assets.