Two checkpoints of nominally the same model produce different images. A pipeline that says which model it uses but not which checkpoint has not said anything reproducible.
For hosted models the equivalent is the version string, and it moves without asking you. A campaign generated across a silent provider update is a campaign with an invisible seam in it.
Record the checkpoint or version in the set specification and in the run log. It is one line and it is the difference between being able to regenerate a frame in six months and not.
Why record the exact checkpoint or model version?
Because a provider can update a hosted model without notice, and a set generated across that boundary will have an invisible seam that nothing in the files explains.
What if the version is not published?
Record the date and time of the run instead, along with anything the provider does expose. It is weaker but it at least brackets the problem.