On mature library code the false-flag rate was 0 of 30 — but polished library internals are the easiest possible case, so the 14% above is the figure to plan around. Both were measured at a 75% threshold; this page flags from 70%, which is lower, so expect at least that many false flags here.

Not yet validated on real student homework, and detection is weaker against some models than others. We cannot predict in advance which human files will read high.