Antennal grooming

Real-lab code (Ozdil et al., Nature Communications 2026, Ramdya lab). It genuinely uses connectome data - the grooming network is built from FlyWire neuron IDs - and the simulated network is a documented neuron model: flyvis GroomingDynamics defines 'passive point neurons with instantaneous graded release synapses' with a per-neuron time constant and a voltage ODE dV/dt = (1/tau)(-V - bias + W*r(t) + u(t) + noise), and this repo's figures are built from the resulting voltage traces. Not L2: the only place a synaptic weight is ever written is offline supervised training in the flyvis fork, where the per-edge syn_strength parameter carries requires_grad=True and is fit by torch.optim.Adam through loss.backward()/optimizer.step(). That is an optimizer's gradient, not a rule grounded in fly biology, and it runs before the simulation, not during it - at simulation time (the Fig. 6 current-injection and activation/silencing screens) the weights are fixed. The rating is a statement about what the model does, not about the quality of the science. Under the scale as it now stands this carries the validation mark: Figure 7 tests a model prediction in live flies with optogenetic activation plus an air puff, measuring foreleg-to-antenna distance per fly against a no-driver negative control. That is a behavioural comparison against real animals, and it sits beside the level rather than above it — which is the whole reason validation was split off the ladder.