In our solver formulation we have an observation model of the form
$$z_i \sim \mathcal{N} \left( z_i; \dot{y}(t_i) - f(y(t_i), t_i), R \right),$$
which, together with the zero-data ${z_i = 0}$ define the information operator of the solver. Currently, we always assume $R=0$ as we are concerned with solving ODEs. But, the algorithm would in principle also work with $R>0$, which could be e.g. used when solving discretized PDEs [1]. So let's add this.
In our solver formulation we have an observation model of the form $$z_i \sim \mathcal{N} \left( z_i; \dot{y}(t_i) - f(y(t_i), t_i), R \right),$$ which, together with the zero-data ${z_i = 0}$ define the information operator of the solver. Currently, we always assume $R=0$ as we are concerned with solving ODEs. But, the algorithm would in principle also work with $R>0$, which could be e.g. used when solving discretized PDEs [1]. So let's add this.
[1] "Probabilistic Numerical Method of Lines for Time-Dependent Partial Differential Equations", Krämer et al, AISTATS (2022)