fly_chess

A documented LIF model — dt 0.5 ms, tau_m 20 ms, v_rest -52 mV, v_thresh -45 mV, 2 ms refractory, all in a LifConfig dataclass — run with shuffle controls, and real MaleCNS v1.0 is used for the 475-cell identity/circuit slice (sugar -> MN9, loom -> DNp01) that the `circuit --source malecns` path drives. Not L2, and the project is explicit about it: 'Graph frozen' is repeated across the codebase and value_head.py's own docstring reads 'Graph stays frozen. Only these weights train.' This is the trained-readout-on-a-frozen-graph pattern the scale is designed to catch, and the authors caught it themselves: the README reports `wiring_helped` is false, that the real graph is silent at 0 Hz after one LIF ply, that a shuffled copy performs the same, and that the chess player is 'the register plus search', not the fly. One caveat a reader should know: the chess measurements run on `build_fixture`, a hand-authored 1,007-cell graph named after real cell types but wired in code — the connectome proper is only exercised by the identity and circuit probes.