Multivariate z-score surveillance across 24 distributors and 6 operating metrics. Each cell is scored against a rolling 8-week trimmed baseline — relative to its own history, not peer-compared. The system surfaces 2 critical, 5 flagged, and 33 watch-list conditions this week.
Each cell = (distributor, metric). Click ◆ cells to drill into the forensic story.
The bulk of cells sit within ±1.5σ as expected under the null; the tail cells are the ones the surveillance grid escalates for review.
Hold has stabilized at 11.4% vs an 8-week trimmed baseline of 7.6%. Onset coincides with the W11 PRF math model rollout — most likely a planted change, not an anomaly.
Confirm with Math team that the W11 model is the designed change. If so, suppress this anomaly from the weekly flag list and re-baseline.
For each (distributor, metric) cell, the current week's value is compared to the trailing 8-week trimmed mean — the highest and lowest observations are dropped before averaging, which makes the baseline robust to single-week shocks. The deviation is divided by the trimmed standard deviation of the same window, yielding a z-score naturally normalized across metrics of different scales.
Three tiers structure the response. Watch (1.5 ≤ |z| < 2.5) accumulates without action. Flag (2.5 ≤ |z| < 3.5) opens a forensic investigation. Critical (|z| ≥ 3.5) escalates to the weekly leadership briefing. Under a standard normal null, about 1.2% of cells reach |z| ≥ 2.5 by chance — roughly 1.8 of 144 cells. This week shows 7, materially above that baseline and consistent with real signal rather than noise.