Calibration after jointly refitting the neural SIR and reporting layers

Evaluate the calibration of the fitted neural SIR report procedure after jointly repeating the neural epidemic-model fit and the reporting-layer fit, rather than holding both fitted components fixed during assessment.

Background

The paper’s weekly-report diagnostic constructs a reference law from an already fitted neural SIR implementation together with a fitted negative-binomial reporting model. The reported comparisons hold the fitted neural and reporting parameters fixed, so they assess score behavior conditional on those fitted components rather than incorporating uncertainty from repeating the full fitting procedure.

The paper establishes calibration results for Monte Carlo comparisons under conditional matching of target and reference score distributions, and it studies fitting effects in separate controlled experiments. However, the complete procedure that jointly refits the neural SIR dynamics and reporting layer before conducting the weekly-report diagnostic is not evaluated. Establishing its calibration would determine whether the stated reference-rank procedure remains reliable when both model-fitting stages are repeated.