Coverage controls changed the four-map Gaussian error-score gain from +0.0024 to -0.00036 nats in a fixed C. elegans ridge test

We tested whether disagreement between reference connectome models helps predict errors in measured neural responses. We used 6,270 WT calcium response means from one C. elegans atlas. Each graph kept all 302 canonical neuron slots and all recorded nonself neuronal edges. Separate ridge models used four mature exports. Nested class-pair folds kept reverse directions and bilateral copies together. Mean prediction spread was 0.001670, or 1.97% of the ensemble squared error of 0.084658. Adding spread to a Gaussian error model with measurement count and predicted response size improved test score by 0.002419 nats per direction. Its descriptive 95% endpoint-weight range was -0.002679 to +0.011122. The 276 directions with both endpoints outside shared map coverage supplied more than the total positive gain. A later control fitted separate error models for zero, one or two endpoints outside common coverage. The added disagreement gain then became -0.000356 nats, with range -0.003215 to +0.002107. Planned partition, transform and map checks also had ranges across zero. These results give no clear added error-score gain under the tested models. The maps mix coverage, source material and weight units. The response export lacks animal and event IDs. The ranges condition on fixed predictions and do not describe new animals. The package preserves complete graph arrays, response moments, all fitted states, source records, figures, an inspection tool and repeat code.

connectome · Caenorhabditis elegans · structure and function · prediction error · anatomical coverage · model disagreement · calcium response · ridge regression