After 1 s of gray input, a 20 ms update lost at least 3,241 voltage directions in each of 50 released fly visual models

We tested whether the 20 ms update in released flyvis 1.2.0 models removes full voltage-state information. We used all 50 models, with their wiring and learned parameters fixed. Each model started at its learned bias and received gray input for one second. If a coordinate has a learned time constant at or below 20 ms and a negative voltage, its direct retention is zero and its rectified output is zero. Changing such a voltage while it stays negative gives the same next full state. This exact test found 3,241 to 10,596 independent lost directions per model, or lower bounds of 7.1% to 23.2% of the 45,669 model coordinates. The median bound was 14.9%. There were 4 to 17 certified coordinates among the 65 central units. A voltage margin of 0.001 gave a median bound of 6,795 directions. Five full-network perturbation tests confirmed the map identity to rounding precision. Smaller Euler steps retained the perturbation and approached the conditional continuous decay of exp(-1) after 20 ms. Source arrays rebuilt exactly, and a clean repeat matched all new numerical results. Under the usual ODE uniqueness condition, distinct full starting states cannot merge in finite time. Thus this update cannot equal such a flow on the same state space near the tested states. The result concerns model state, not measured fly physiology, visual task loss or useful memory. The package preserves the proof, code, source provenance, arrays, results and figures.

connectome · fly vision · numerical integration · state retention · Euler method · reproducibility