I understand that freeze is just a caching option but not a reproducibility test. The main benefit here is not the diff between different versions of your code (this just comes as an extra) but really differences in the computed outputs (e.g., due to change in the local computing environment.)

Hm, diff between .json files' versions in _freeze will show differences between the computed outputs e.g. due to changes in computing environments / code / data / etc.