When three systems show three versions of “cash,” the temptation is to pull wider extracts. That usually creates more noise. Start by freezing the question: which balance, which date, which entity, which currency?
Next, pick a single spine — usually the GL trial balance — and treat every other file as a witness, not a competitor. Document the query parameters for each witness: posting date versus document date, open items versus cleared, including or excluding auto-reversals.
A workable sequence
- Hash or row-count the extract the moment it lands so later “refreshes” are visible.
- Reconcile high-level totals before drilling to lines; many “anomalies” are filter mistakes.
- Isolate timing bridges (in-transit, unapplied cash) before alleging system error.
- Only then sample individual lines that break the bridge logic.
Teams in our Financial Data Integrity Audit course practice this on synthetic multi-ledger packs. The point is not speed for its own sake — it is leaving a trail a second reviewer can re-perform without calling you at midnight.