Many consequential incidents are not schema failures. Every record is well formed, yet systems disagree about inventory, demand, order state, spend, or time.
What made it hard
Each source may be internally consistent. The conflict becomes visible only when claims are compared in business and temporal context. Choosing one source too early can erase the evidence needed to understand the failure.
How I approached it
I used layered controls: technical audits, business-state balancing, ratio and distribution checks, temporal validation, error history, proactive alerts, and restart-aware recovery. The system preserved disagreement as evidence instead of hiding it behind a single status.
Selected impact
The approach made cross-system disagreement observable through explicit balancing, exception history, and restart-aware controls. Public descriptions intentionally omit internal business-performance figures.
Lesson I still use
Data quality is not a final report. It is part of control flow, ownership, and recovery.