Scrollfly

L2 is ruled out by the authors themselves — they state the readout was trained on 500 videos and that 'The full connectome remains frozen', and they report the honest negative result that the readout does not beat the constant baseline on their 50 creator-disjoint clips. But L0 vs L1 cannot be settled from this repo, because the simulator is not in it. The published site is a replay: it loads pre-recorded per-neuron firing-rate bins from release assets and paints them onto MaleCNS SWC skeletons, and the environment page is explicitly 'interactive MaleCNS anatomy without simulated neural activity' driven by a hand-written pixel-classifying reflex that its own header calls 'a visual reflex, not a trained agent'. The only path to a live brain is an HTTP client POSTing to a /infer service on 127.0.0.1:8877 that the README says 'needs a separate compute service and is not exposed'. Releases contain only the pages bundle and the Chrome companion. To settle L0 vs L1 a reader would need the source of that inference service — specifically whether it integrates membrane voltages under a stated neuron model, or maps connectome activity through a static linear map.