FlyScroll

A compiled all-edge LIF simulation of the MaleCNS graph (Shiu-like constants: -52 mV rest, -45 mV threshold, 20 ms membrane, 5 ms synaptic, 2.2 ms refractory, 1.8 ms delay) in which sustained Kenyon-cell firing depresses the existing KC synapses onto the novelty MBONs (MBON16, MBON17, MBON28 - the alpha'3 novelty compartment, cited in the source to Hattori et al. 2017), with slow homeostatic recovery toward baseline. The depression is written into the brain's own edge weight array from the measured per-bin KC spike rates, bounded to 0.25x-1.0x baseline, and that array is the one the LIF kernel propagates through. Learning is on by default (--no-learning is opt-out), and there is no optimizer, trained readout or offline-fitted weight file anywhere in the project - the scroll decision is read straight from population novelty. Unlike its siblings this rule is activity-driven rather than dopamine-gated: PPL104 (PPL1-a'3) is named in the source as the cell that gates this compartment's plasticity, but it never enters the update. Two corrections a reader should have. The surprise-gated dishabituation method exists in HabituationRule but is dead outside the unit tests: at a reel boundary the session calls only begin_reel and records the policy as 'continuous_stimulus_specific_homeostasis', deliberately leaving recovery to the continuous homeostatic term rather than a transition reset. And the source justifies MBON16 and MBON17 as alpha'3 from the MaleCNS instance names but gives no such justification for MBON28. Not validated: the repo publishes no comparison of its habituation dynamics against real fly novelty-MBON measurements; there is no validation document at all.