A fixed unsigned C. elegans connectome score raised calcium duration prediction error by 0.65% after local controls for excluded animals and cell classes

Can a fixed network model predict how long neural activity persists? We used all 302 neuron slots in the Cook adult C. elegans reference. Function data had 1,776 named neuron observations from 21 freely moving worms. The model had positive normalized weights, equal leak, and independent equal noise inputs. Its score was the exact integral of each model autocorrelation. We tested whether that score adds to local wiring totals, degrees, broad neuron groups, and reference position. Every prediction excluded its animal and cell class fold from fitting. Adding the score raised mean squared log-duration error by 0.65%. The 95% recording resampling range was 0.41% to 0.91%, with predictions fixed. Error rose in 19 of 21 animals. All planned variants had negative mean gains. In added tests, a simpler model gained 3.43% against one mean only by reversing the network duration order. Broad group controls reduced that gain to 0.044%. A separate known-neuron task used other animals and reduced error by 32.7%. Thus, observed duration differences were reproducible, while this fixed network score gave no added gain in the main conditional test. The result limits this model and readout. It does not establish that wiring or recurrent circuits lack a role in persistence. Nuclear calcium, behavior removal, different source animals, extrapolated anatomy, and fixed model assumptions limit inference. The package preserves code, numerical summaries, exact predictions, source versions and hashes, figures, and repeat steps. Independent source readers and numerical methods checked the main results. A full repeat matched 39 numerical files exactly.

connectome · C. elegans · neural dynamics · calcium imaging · autocorrelation · neural timescales · linear systems · reproducible research