Something changed.
Know where to look.

The first 32 observations establish a baseline for 48 variables: temperature, pressure, humidity, and wind across 12 simulated sensor groups. Each new 16-observation window is compared with that baseline, starting at observation 48.

The mean component is the root-mean-square change in all 48 standardized channel averages. The relationship component measures changes in the 12 within-group temperature/humidity correlations. The combined index is the square root of the sum of their squares. This selected-pair summary does not capture every possible high-dimensional change.

In the relationship scenario, the generating relationship changes at observation 65 while individual generating means and scales stay the same. Sample averages still fluctuate. The drift scenario changes several averages; the stable scenario adds no intentional change.

Three consecutive windows above the chosen threshold record a review trigger. High, balanced, and low use 0.45, 0.60, and 0.85 index units. These are illustrative settings, not calibrated probabilities or significance levels. Only observations up to the timeline position are analyzed. The datasets are fixed synthetic examples, not validation results.

The first 24 observations establish a baseline mean and standard deviation. The detector then compares each 8-observation moving average with that baseline. Three consecutive averages beyond the selected threshold trigger a review marker.

High sensitivity uses 1.4 baseline standard deviations, balanced uses 2.2, and low uses 3.2. Higher sensitivity reacts to smaller changes. This simplified example does not model real-world false-positive rates, dependencies, or changing baselines.

Only observations up to the selected timeline position are analyzed. The reported deviation is the current moving average's distance from the baseline, measured in baseline standard deviations.