Stress-Testing Local Neural-Dynamics Equivalence in a Connectome-Constrained Visual System

Abstract

Connectomes constrain which neurons can interact, but neural computation also depends on dynamics that connectomic measurements do not determine. This study asks a compositional question: if an alternative neuron model is locally equivalent to a source model on held-out, source-clamped trajectories, does that equivalence survive when a cell-type block is substituted inside a recurrent connectome-constrained network? I implemented cell-type-masked dynamics substitution in a public model of the Drosophila visual system containing 45,669 neurons and 1,513,231 edges. Three alternative dynamics families were calibrated under a frozen 1% local derivative margin and tested in Mi1, T4a, and T5a blocks across 24 moving-edge trajectories, 25–100% replacement, two pretrained checkpoints, an unchanged optic-flow decoder, activity-impulse interventions, and a halved integration step. Exact shams, topology digests, and deliberately wrong controls audited the measurement path.Local equivalence was neither automatic nor uniform: a two-state substitute failed in T5a and in one Mi1 checkpoint. Among locally eligible cells, however, all 45 replacement-ladder evaluations preserved ordinary block responses and frozen-decoder outputs within the 1% margins; no ordinary composition failure occurred. A more sensitive causal gate found a replicated Mi1 timescale discrepancy: decoder-effect normalized root-mean-square error was 1.897–2.754% across two checkpoints and two step sizes while effect-vector cosines were approximately one. This supports a small gain deviation, not a change in causal direction. A preregistered follow-up attempted to compare dynamics families at matched local error. It stopped before confirmatory holdout exposure because no shunt candidate both passed all local gates and matched the timescale error within the frozen interval.The resulting claim is bounded. In this model, under these stimuli, interfaces, and margins, locally eligible dynamics substitutions were robust for ordinary responses and decoded output, while perturbation effects exposed a narrow gain sensitivity that did not survive the study's matched-control advance criterion. The study provides a staged falsification protocol and a publishable null boundary; it does not establish connectome sufficiency, biological replaceability, consciousness preservation, or a general composition theorem.