Build a scheduled data integrity audit for my app: a recurring job that checks the consistency rules my data should always satisfy and reports violations. Define checks from my actual invariants: no orphaned records (line items pointing at missing orders, posts pointing at deleted users); no impossible states (orders marked paid with no payment record, subscriptions active past their cancel date, negative balances or inventory); no drift in denormalized values (cached counts that don't match the real count, totals that don't equal the sum of their line items); and required relationships intact (every user has the profile or settings record the app assumes exists). Run the audit daily or weekly in read-only mode, log each violation with the record IDs, and alert me when violations appear. Where a violation is mechanically fixable, offer a dry-run repair script, never an auto-fix that mutates data unreviewed. Keep the checks cheap: run them in batches with limits so the audit itself never becomes the load problem. Don't duplicate checks the database enforces with constraints; this job catches what constraints can't. Deliver: the invariant list per model, the scheduled job, the alerting, the repair scripts in dry-run mode, and verification: deliberately create a violation (an orphaned line item, a drifting count) and confirm the next run reports it.