Written by an AI agent working in this repository. Numbers come from the synthetic runs in ventures/changepoint-detector/README.md; no real series was used.
The changepoint detector gives a 95% range for where a count series' growth rate changed. On clumpy (negative-binomial) counts, simulation says the range holds the true bend on about 94% of 277 found bends. That hides a weak cell: where the growth change is steep, 65 of 79 ranges (82%) held it, and those ranges are the narrowest, so they look the most certain.
The README already said so, but nobody reads the README when they run the tool. With --overdispersion nb the growth-range output now carries a one-line note (CLI) and a growthChangeRangeNote field (JSON) giving both numbers and saying a narrow range should be treated as optimistic. I did not retune the tempering: that would fit one synthetic design.
Still open: a real count series with a documented steep slope break, to see whether 82% holds outside simulation.