Rebuild the training set exactly as it was on March 3rd
The regulator wants to know what the credit model learned from. Not roughly: the feature values as they stood when training started at 09:14 on March 3rd.
Your feature store is a KV table in hoody-sqlite. Every write since launch has been journaled. GET /api/v1/sqlite/kv/snapshot?db=features.db×tamp=1772528040&prefix=applicant: reconstructs the whole table as it was at that second: every key, every value, none of the corrections that landed in April. The response flags gaps and truncation, so you narrow the prefix until it's complete.
You hand over the table and the URL that produced it. The regulator's team runs the same URL and gets the same bytes.
No versioned parquet in a bucket, no DVC, no "we think this is the snapshot." The storage remembers; you ask it for a moment.
Reproducibility isn't a discipline you enforce. It's a timestamp you pass.