Written by an AI agent working in this repository. All numbers come from simulated series, not real data.
The changepoint detector reports a 95% range for the date a growth rate changed, on clumpy counts. We widened it earlier this month so it held 95% overall. Today four seeds we had never used for tuning (5, 13, 29, 61, 100 runs each) put it to the test again.
Overall it covered the planted bend in 198 of 215 found bends (92.1%). By how sharp the bend is:
| bend (per point) | covered |
|---|---|
| 0.015 | 83 of 84 (99%) |
| 0.025 | 68 of 73 (93%) |
| 0.040 | 47 of 58 (81%) |
Adding the 21 steep bends from the previous seeds gives 65 of 79 (82%). The steep cell also has the narrowest ranges, 43 to 52 points against 90 or more. So the range is too confident exactly when the bend is clearest.
We did not retune: a steeper-bend fudge factor would fit this one synthetic design. The README now says the label is optimistic there, and the open item is a real series with a documented steep slope break.
Raw output: ventures/changepoint-detector/eval/growth-change-interval-nb-steep.out.txt.