How we rate
One level per project, plus a mark for whether anyone checked.The scale measures one thing: how much of the fly's biology a project actually simulates. It is not a measure of effort, polish, cleverness or how fun the demo is. A beautiful browser game can sit at L0 and a rough research script at L2 — that is the scale working, not failing.
The connectome is used as fixed wiring. Activity is mapped straight to controls. Nothing changes over time.
Evidence needed: the mapping from neural activity to output, in the source or the docs.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Evidence needed: the neuron model and its parameters, written down somewhere a reader can check.
Synapses change under a modeled rule grounded in fly biology — for example dopamine-gated change on Kenyon-cell → MBON synapses in the mushroom body. Learning begins here.
Evidence needed: the learning rule itself — which synapses change, driven by what signal, on what timescale. Generic Hebbian or STDP applied across the whole graph is real plasticity and still not this: the rule has to be the fly’s, naming cells or compartments, gated by a modelled neuromodulator.
Behaviour or activity has been compared against measurements from real flies, and the comparison is published. It can accompany any level — a simple model that was checked beats an elaborate one that was not.
Evidence needed: a published comparison against measurements from real flies, with the measure stated.
The rules
A level needs evidence, not a vibe
Every level published here ships with the claims that earned it and a link per claim. If nobody has read a project closely enough to file that evidence, it stays unrated. An honest gap beats a confident guess, and most of this index is currently a gap.
Validation is a mark, not a rung
The scale used to end at a fourth level, “validated”, stacked on top of plasticity. A published VNC walking model broke it: its prediction was confirmed by optogenetics in live flies, with the same rhythmicity measure applied to real and simulated neurons — and it has no plasticity at all, so a cumulative scale filed the best-checked work in the index alongside a sandbox.
So the two questions are kept apart. The level says how much biology is simulated. The ✓ says somebody checked it against a real fly. A simple model that was checked is worth more than an elaborate one that was not, and the scale should be able to say so.
One level, the highest one fully met
Levels are cumulative. A project sits at the highest level whose evidence bar it clears — not the highest one it gestures at. Plasticity code that exists but is documented as not working does not clear L2 on its own; plasticity that runs and is described does, even when the project reports that nothing was learned.
Three things that never move a rating
- Stars, forks, or how widely a project was shared.
- How good the demo looks, or whether it plays Doom at all.
- Motion. A fly that walks is running a body model, not a learning brain.
A trained readout is not a learning fly
Many projects freeze the connectome graph and train a small artificial readout on top with reinforcement learning or CEM. That is a legitimate and often carefully done experiment — and it is not fly plasticity. It does not reach L2 unless synapses in the modeled fly circuit change under a rule grounded in fly biology.
Who reads these, and how
Worth stating plainly, because the site asks projects to be precise about their own methods: these ratings are produced by reading source code with an AI — clone the repository, find the line where a weight is written, decide what changes it. They are not the verdict of a fly neuroscientist, and nobody here has re-run these projects.
That is why the evidence rule is not decoration. Every claim carries a link pinned to a commit, so the reasoning is checkable by anyone who disagrees, and a script verifies that every one of those links still resolves. The rating is an argument with its working shown — judge it on the links, not on who wrote it.
Status is part of the rating
- Draft — one reading, unchecked. Shown with a [draft rating] mark everywhere it appears.
- Reviewed — a second, independent reading checked the evidence against the source and agreed. The second pass is not a formality: it overturned five proposed L2s, four of them for running a generic learning rule rather than a fly one.
- Disputed — an open, public disagreement. The level stays visible while it is argued; it is not quietly changed.
A human reading that contradicts any of these outranks it. That is what the dispute link on every project page is for, and it is the most useful thing anyone can send.
Disputes are public
Every project page carries a dispute link. A dispute needs the same thing a rating does: a claim and a link. How to file one →
Know one we missed?
Disagree with a rating?
Submissions and disputes are public. Every level ships with the evidence that earned it.