With motor labels and thresholds fixed, one neuron label change altered at most two of 106 feeding calls in a positive three-step FlyWire rule

We tested one excitatory or inhibitory neuron label error in a fixed positive path rule for fly feeding responses. We used all 127,400 neurons and 52,793,639 chemical synapses in the original FlyWire v630 model graph. The rule adds input-normalized path weights over one to three steps to the two MN9 motor neurons. The screen has 106 stimulated cell types and 14 measured positives. The original rule gave 99 correct calls. We held its decision thresholds fixed and reversed each of 127,013 neuron labels outside the motor pair. Each reversal changed all outputs of that neuron together. No reversal changed more than two calls. The correct count stayed between 97 and 100. Only 39 reversals changed a call. Across different possible errors, 20 calls could change. All three two-call changes came from shared input roots in the Nori and Oink input sets. When each test's own input labels stayed fixed, only Clavicle and Rattle could lose a positive call through a relay label change. We give an exact matrix update method and its proof. Small graph checks and 260 full graph checks agreed with direct calculation. A fresh graph rebuild gave exact agreement for 34 result tables and audit files. The result applies to this fixed path rule and one label error. It does not establish robustness of living flies, spiking models, or several simultaneous errors. The package preserves code, source hashes, all score changes, thresholds, witnesses, figures, and repeat instructions.

connectome · FlyWire v630 · structure and function · feeding behavior · transmitter labels · model robustness · single neuron error · exact matrix update · reproducibility