A shared complex map reduced left-bend coherence error by 7.5% in 35 inputs to one leech motor neuron

A shared gain, phase change, and offset reduced prediction error by 7.5% relative to the target-trial mean in 35 identified inputs to leech motor neuron DE-3R. We used three left local-bend trials and anatomy from the same ganglion. Each prediction excluded its target input from the fit. A whole-input identity reference gave p = 0.0003 from 9,999 draws. Error reductions for the three trial pairs were 24.1%, 0.9%, and -0.6%. The result was therefore modest and uneven. In two swim trials, the same method gave an 86.1% reduction. Strict confidence scores gave similar left-bend reduction, while cross-behavior transfer was weak. Population calibration recovered some information missed by a raw earlier-value prediction, but left most target-mean error. One animal, few trials, shared circuit inputs, uncertain matches, and absent raw trace errors limit inference. The map does not identify a causal wiring mechanism. We provide code, source hashes, retrieval instructions, predictions, label draws, figures, and independent calculation checks.

connectomics · leech · coherence · population alignment · repeated trials · structure and function · prediction