The window sizing tradeoff is the part I've struggled with too. On a filings dataset I work on, corrections can show up months later when a company amends a report, so a fixed lookback either misses them or rescans far too much. What helped was keying reconciliation off what actually changed upstream instead of a date window. Curious whether IFCO's late events had a hard upper bound on lateness, or if the window was picked from observed data.