Currently, early measurements (typically fourth, relative to third) can result in surprisingly high loss count. This makes the next measurements to be done deep in zero loss region, and the algorithm would need too much time to converge back to the critical region.

Early attempts at discarding or suppressing outlier measurements lead to highly unstable behavior, so for now the returned value has to take all measurements into account.

But it might be possible to use a kind of suppressed (thus wrong) computation just for the purpose of next target TR for measurement. Multiple unsurprising measurements would still be needed for the final bounds to converge, but measuring in the critical region should make the convergence faster.

Possibly, tweaking time parameters for optimal speed can be part of this task.