submit a result

Pick parameters. We run the brain.

Nothing to install and nothing to download. This form writes a small config file; GitHub opens it as a pull request (it makes a fork for you if you need one); our CI runs the full 12-task suite on the real FlyWire connectome with your settings and comments the scores on the PR. When it's merged, your row appears on the leaderboard marked verified.

labelshown on the leaderboard; make it yourswhat you changedone line; this is what other people readconflict of interestwho wrote the model, whether you wrote tasks it targets, what you stand to gain; none is an answermodelbuilt-in simulators only; custom code goes through a normal PR

gainvolume knob on every synapse. 1.0 is the 2024 paper; 0.45 is the robust window on today's data

seedsrandom repeats; a check only passes if it holds on every seed. 3 is the standard

the file GitHub will open for you

label: …
note: …
connectome: flywire783
seeds: 3
params:
  gain: 0.45
simulator: flybench.sim:LIFSimulator
conflict_of_interest: "none"

· label: 1–60 characters; letters, digits, spaces and . _ , ( ) + / -
· note: say what you changed, in one line

Watch this gain in the explorer first →Open pull request on GitHub →

GitHub opens with the file pre-filled (sign in if asked; it forks the repo for you).
Click Propose changes, then Create pull request. Don't edit anything.
Within ~10 minutes a bot comments the scores on your PR.
When a maintainer merges it, your row is live on the leaderboard.

What a higher score means, and what it doesn't

A score counts how many of the listed fly behaviours the model reproduces. It does not say the model is closer to a real fly's biology. A model can pass more tasks by adding a mechanism flies don't have, or by adding a real one with the wrong numbers. Treat every row as a hypothesis with its own evidence: the note, the constants, and which tasks moved. The benchmark is a way to argue with numbers, not a verdict.