Chart 1 — rolling mean underwriting-score by applicant group

The chart your Chief Underwriting Officer should look at first. Where the silent I3 divergence begins, in one visual.

What you are looking at

• X-axis: Day 1 through Day 90 of the audit period
• Y-axis: 7-day rolling mean of the AI's underwriting score (0-100 scale)
• Each line: one applicant group (I1 through I5, with I3 = age 25-45 / mid credit / mixed urban-lower-income zip)
• Baseline window: the first 30 days establishes what "normal" looks like per applicant group

What the eye sees

Four applicant groups stay roughly parallel across the 90 days. One applicant group — I3 — rises. The rise begins around Day 30, and it does not recover. The line separates from the pack, climbing toward the decline / refer-SIU threshold at score 67.

This is the pattern an underwriter-in-the-room instinctively recognizes: something in the I3 stream changed, and the AI is scoring it as higher risk than baseline. The 51 percentage-point lane-shift follow-on is the operational consequence.

What the sensor fires on

The sensor is not looking at the eye's pattern. It runs four quantitative thresholds against each 7-day window per applicant group:

• Mean-shift: is the 7-day mean > 1.5 standard deviations from baseline mean?
• Variance-ratio: is the 7-day variance > 1.5x baseline variance?
• KL-divergence: is the shape divergence between baseline and current distribution > 0.2?
• Group-differential: is the max cross-group shift > 12%?

Day 44 fires I3 KL-divergence 4.378 (roughly 22x the threshold). The pattern the eye sees is the pattern the sensor quantifies.

What regulators care about

The chart also tells the counterfactual story: had this monitoring been in place operationally, the drift window would have been ~7-14 days rather than 60+.

The next chart

The rolling-mean chart shows when. The baseline-vs-drifted chart shows how the I3 distribution shape shifted.

Chart 2 — baseline vs recent distribution →