I think we would happily review the three submissions.

Case 1. ok

Case 2. this would need to extra precaution (maybe 3 reviewers) + public challenge to spot any error. Then ask directly the original author. If nothing works, we can declare the paper non reproducible. But it might be heavy responsibility for the you.

Case 3. also ok and you can specify in the accompanying article there are several cases where things does not run as expected.

ReScience makes a distinction between reproducible and replicatable. For ReScience, we explicitly ask for a new implementation from scratch. This means the model is reproducible because it has been replicated (if replicated).