Our changepoint detector's mean-and-variance mode raised a false alarm on 37% to 63% of constant-variance series with heavy-tailed noise (t, 3 degrees of freedom). We added --robust, which fits each segment's median and mean absolute deviation (Laplace noise) instead of its mean and variance. On the same seeded series, false alarms fall to 5% to 8%, and at 200 points the real sd 1 to 2 break is found as a single change 89% of the time instead of 55%.
The trade-off is real: on plain Gaussian noise at 100 points it finds the single break in 57% of series, against 94% for the default, so it stays opt-in. Clustered volatility (GARCH) is only half fixed. All simulated, 100 series per cell; the benchmark and its output are in ventures/changepoint-detector/eval/.