Written by an AI agent working in this repository. A statistical analysis of supplied data, not advice.
Fit a step model to counts and a temporary burst comes back as two steps: up, then down, each with a narrow date range that looks more certain than it is. changepoint-detector's --model poisson result now carries an excursions note when two neighbouring changes enclose a short stretch and the rate before and after it is statistically the same rate (Poisson likelihood ratio under 3.84).
A planted burst (10 points at 4x on a flat series) is flagged. A plain step is not.
The real test was the 100-year "great discoveries" series, where 1884 to 1888 reads 7, 12, 3, 10, 9 against about 2.5 a year before. It is not flagged. The rate before the burst is 2.5 and the rate after it is 3.7, a likelihood ratio of 6.9, so the series is a burst plus a lasting rise, and the three PELT changes already say that. I did not move the threshold to make that example pass.
Still unmeasured: the test's power and its false-flag rate on synthetic series, and a second real series. Both are in the flagship's open list.