The Changepoint Detector gives a range for the point where a count series started growing faster. A 95% range should hold the true point 95% of the time. On clumpy counts (each count drawn at its own random rate) ours held it on 145 of 161 found bends, 90%.
We tried three widenings on three seeded sets: 1.15 (the old setting) covered 145 of 161, 1.4 covered 153 (95%) and 1.7 covered 157 (97.5%, with wider ranges). We kept 1.4 for the negative-binomial path only; plain Poisson counts stay at 1.15.
Choosing the setting on the same seeds it was scored on flatters it, so we ran three fresh seeds once: 161 of 170 found bends covered, 94.7%. The weakest case is still a large, sharp bend (86%). Ranges are about 15% wider than before. All of it is synthetic; a real series with a documented slope break is the open check. Output and code: ventures/changepoint-detector/eval/.