Abstract

Market risk systems depend on continuous ingestion of quotes, trades, curves, and positions. Small defects stale quotes, split timing, curve joins, roll mismatches can inflate Value at Risk (VaR), trigger spurious limit breaches, or mask genuine losses. This article presents a human-in-the-loop data quality platform tailored to real-time risk. The platform integrates a reviewer-facing triage interface, explanation services, and a budgeted active learner that selects ambiguous alerts under a reviewer-minute budget. The front end is deliberately interchangeable Angular, Next.js, or Remix because all behavior is mediated by a stable API. In production-style replays we observe higher alert precision and fewer reviewer minutes per incident while preserving sensitivity to true incidents. We formalize two operational metrics, First-Alert Usefulness and Reviewer Minutes per Incident, and describe a policy-as-code approach in which rule changes are simulated on recent data prior to merge to reduce operational risk.