fly-explorer

A genuine leaky integrate-and-fire simulation over real FlyWire v783 data (the 32 MB export is committed): membrane voltage integrates toward rest with a synaptic conductance term, crosses threshold, resets, and honours a refractory period and a 1.8 ms conduction delay, with the five constants taken from Shiu et al. 2024. That is a documented neuron model, so L1. Not L2: the CSR weight array is read-only - there is no optimiser, no readout fit, no plasticity term of any kind in the repo, and the app itself says so. No validation mark, and it is the closest near-miss on the site. The project publishes a leaderboard scoring itself on 24 tasks against quantitative recordings from real flies — DA1 projection-neuron rate against Olsen, Bhandawat & Wilson 2010; giant-fibre spikes per loom against von Reyn 2014 and Ache 2019 — and reports both numbers, failing several of them openly. That is the honest kind of benchmark. The mark is nevertheless reserved for a comparison published through peer review, because a benchmark a project scores itself on is too easy to make flattering, and the filter is what keeps the mark meaning something.