Test Database Migration Rollbacks Against a Disposable Copy Before Deploying

AI-generated database migrations often produce correct forward scripts but flawed rollback scripts, since models must infer a table's previous state without access to the baseline schema. Running the rollback path against a throwaway database before deployment can catch these errors early, before they compound a production incident. The recommended approach captures the schema before applying the up script, runs both up and down scripts on a disposable database, then compares the resulting schema to the original. A lightweight Python harness using SQLite can automate this check on every pull request, making rollback verification a repeatable property of the migration rather than an optional review step. If the down script fails, the mismatch output can be fed back to the model to refine only the rollback file, improving accuracy without rewriting the entire migration.