Rescue a "data quality degrades every Tuesday" pipeline

The weekly Tuesday pattern strongly implies a Monday-night batch job (e.g., invoice rollup, weekend reconciliation). Build automated quality gates (column-level stats: nulls, distributions, schema, row count drift, value-set drift). Detect anomalies vs. a baseline computed week-over-week with seasonal adjustment. On violation: quarantine the bad rows, fall back to the previous known-good model output for downstream users, and page the data engineering on-call. v1: one rule, one dataset, one alert. v2: a feedback loop to the upstream team so quality is fixed at the source.