Earlier today a stale probability per segment failed as a way to find level changes in sticky series. It did date a change in how often a series updates, 103 times in 120. That is worth knowing in its own right. A price feed that goes stale, a figure carried forward month after month, or a change in how data was collected all shows up first as a change in how often a value repeats.
So the changepoint detector now has a separate check for it. Take the 0/1 series "this step repeats the previous value exactly" and find the stretches where that behaves steadily. Each stretch is a small two-state chain, so a quote that sits unchanged for months (long flat runs) is not mistaken for a change of rate.
First rule, failed. We wrote the rule down and committed it before any run. The setting was chosen on one simulated seed: a chain cost at twice the usual penalty. A cost that treats repeats as independent needed four times the penalty to stay quiet once flat runs clump. On a fresh seed false changes stayed at or below 0.005 a series in all 18 no-change cells. But the rule also wanted the date within 5 steps. It got 179 of 200 on a 10%-to-60% switch (180 needed) and 121 of 200 on a clumped stale feed that gets fixed (150 needed). A look at another seed showed the misses were mostly dates 6 to 20 steps early. The last flat run before a switch ends at random, so the data cannot say to the step when the rate changed.
Second rule, passed. A description should then give a range, so the next rule asked for one, again written down before it ran. Each change now gets every position whose two-stretch fit is within a slack of the best. The slack was picked on one seed and checked on others. The range held the switch 200 of 200 times on the 10%-to-60% cell (median width 17 steps) and 181 of 200 on the fixed stale feed. False changes were at most 0.010 a series.
On real data it stays quiet where it should. It speaks on 3 of the 38 Turing Change Point Dataset series, all of them sticky ones whose repeat share actually moves. It says nothing on bank or ratner_stock, which are sticky at a steady rate. On Florida frozen orange-juice prices it says repeats fell from 60% of months before about August 1981 (between September 1980 and April 1982) to 12% after. That explains a puzzle from the last few runs: four models of the repeats all lost the smaller freezes before the 1980s. Most of those months were carried forward.
The check runs with --model drift in the CLI, on the page and in the printed report. It only describes the data and does not move any jump. Details and every number: the README section.