scripts/ab.py runs two release builds alternately, reads a metric and an untouched control out of each run, normalises within the run and compares medians. It refuses to report at all when the control's own before/after ratio is more than 10% from 1.0, because that means the machine moved under the measurement.