PyRenew supports scalar ascertainment, but its APIs do not yet define end-to-end behavior for time-varying ascertainment.
Update the model contract as follows:
MultiSignalModel passes n_timepoints and first_day_dow to every registered AscertainmentModel.sample().- Fixed ascertainment models may ignore this context and continue returning scalars.
- Time-varying models return one-dimensional trajectories with shape
(n_timepoints,).
- Ascertainment models declare whether they require a calendar anchor. Require
obs_start_date only when a registered component does.
CountObservation applies ascertainment after clinical-delay convolution, giving time-varying rates observation-date indexing:
potential_counts = convolve(infections, delay_pmf, p_observed=1.0)
predicted_counts = potential_counts * ascertainment_rate
Keep compute_delay_ascertained_incidence() unchanged.
- Existing scalar ascertainment models must retain their current NumPyro sites, output shapes, and model behavior.
- Add tests for scalar equivalence, full-axis trajectory validation, calendar-aligned sampling, and observation-date ascertainment.
PyRenew supports scalar ascertainment, but its APIs do not yet define end-to-end behavior for time-varying ascertainment.
Update the model contract as follows:
MultiSignalModelpassesn_timepointsandfirst_day_dowto every registeredAscertainmentModel.sample().(n_timepoints,).obs_start_dateonly when a registered component does.CountObservationapplies ascertainment after clinical-delay convolution, giving time-varying rates observation-date indexing:Keep
compute_delay_ascertained_incidence()unchanged.