fruit-fly-fund

Synaptic efficacies inside the modeled fly circuit really do change during a run. A dopamine-gated, baseline-centred anti-Hebbian eligibility rule (adapted from Huang/Luo et al. 2024) integrates Kenyon cell and DAN firing-rate traces each 10 ms bin and the resulting deviations are written straight back into the simulation's synaptic weight array, at genuine MaleCNS v1.0 KC-to-MBON07 and KC-to-MBON11 connections, gated by the real PAM11 reward and PPL101 aversive dopaminergic cells - and that same weight array is what the C++ kernel reads to accumulate conductance. This is L2 for the presence of a modeled synaptic change rule, not for demonstrated learning: the project's own headline result is that memory-on lost to memory-off and both lost to buy-and-hold, and the held-out exam showed the advantage did not generalise. The learning rule is vendored unmodified from upstream Stonkfly; this repo's contribution is the frozen-versus-plastic comparison and the exam/reset/decoy controls. No validation mark: nothing here is compared against real fly measurements - the circuit report literally records "validated": false, and the rule's own docstring calls applying it to this male reconstruction a hypothesis. Note also that a separate logistic readout fitted on the fly's own spikes exists in flyvsly/readout.py, which would be the usual L1 trap, but the L2 verdict does not rest on it and the project itself reports that readout as no better than its base rate.