Wow, maybe something to do with the std dev being all over the map?

Prasad, can you take a look (maybe pairing with Dave) since you're the triage oncall this week?

It's that multiple bisect is not implemented.
If the algorithm matched what's written in the "Bisect Confidence" doc, the algorithm would've bisected on both an improvement between 409210 and 409417, and a regression between 409417 and 409624.

There's an improvement between 409210 and 409417 because the std dev at 409417 is 0, so any deviation from that value would indicate a regression or improvement.

(Because multiple bisect is not implemented, it just picks the first change, the improvement, to bisect on. This is not correct.)

OK, so you're saying that there are:

1. improvement from r409210 to r409417
2. regression  from r409417 to r409624

What I don't understand is that when you give it the initial reference range, where r409210 is "good" and r409624 is "bad", how come it decides that the middle revision (r409417) is also "bad" (even though it is even better than the "good" revision)?