Data Quality Report: EventCo Monthly Ingestion

Generated by agents/ingestion_agent.py. This agent never reads data/ground_truth.md. Every fix and flag below is derived from generalizable heuristics (fuzzy matching, time-series interpolation, budget-relative outlier detection), the same way it would run against real, unseen monthly data.

• Rows in raw file: 124
• Rows after de-duplication: 120

1. BU/category name typos corrected (fuzzy match against canonical names)

2. Duplicate rows removed (same month + BU entered twice)

3. Missing Opex values imputed

Strategy: linear interpolation across each BU's own monthly time series for that column (interior gaps); for gaps at the very start/end of the series, the nearest available value is carried in, since linear interpolation cannot extrapolate beyond observed data. Opex budgets and actuals move smoothly month to month for a given BU, so a BU's own recent trend is a more defensible estimate than a company-wide average or a hard zero. No rows were dropped.

4. Currency-formatted text amounts parsed to numeric

5. Date formats standardized

Raw month values mixed ISO (YYYY-MM-DD), US (MM/DD/YYYY), long-form ("Month YYYY"), and year/month (YYYY/MM) styles. Sample of distinct raw values seen: 01/01/2024, 01/01/2025, 01/01/2026, 02/01/2024, 02/01/2026, 03/01/2024, 03/01/2025, 03/01/2026. All parsed successfully via flexible date inference and standardized to ISO (YYYY-MM-DD, first of month) in the cleaned file. Unparseable values: 0.

6. Flagged values: likely data entry errors (high confidence)

Rule: a line item actual coming in above 4x or below 0.25x its own budget is treated as a probable data entry error rather than a real business event. A single BU/month is not operationally capable of swinging that far from plan without a corresponding, equally extreme movement in related lines (COGS, payroll), which these do not show. These rows were not auto-corrected; they are flagged here for manual verification before any variance analysis uses them.

7. Flagged values: notable variances (informational, not errors)

Rule: IQR outlier (1.5x IQR fence) on that BU/column's own historical actual-vs-budget ratio distribution, excluding anything already flagged above as a likely data error. These look like real, large-but-plausible business swings, well under the data-error magnitude threshold above. They are left untouched in the cleaned data and forwarded as-is for the Variance & Root-Cause Agent to explain, not treated as data quality issues.