Running the same text twice will not return the same result. The rewriter samples from a range of possibilities rather than applying fixed rules, so two runs of one paragraph produce two legitimate versions.
This is worth knowing before you conclude that something changed. If a second run reads worse than the first, that is variance rather than a regression, and running again will usually produce something closer to what you wanted.
It also means you can use re-running deliberately. If one sentence in an otherwise good result reads awkwardly, running that sentence alone often fixes it for a handful of words rather than reprocessing the whole document.
Detection results are more stable but not perfectly so, because a sentence near a boundary can land either side of it on different runs. A sentence that flips between mixed and AI is telling you it is borderline, which is information in itself.
Longer inputs tend to produce more consistent output than short ones. A single sentence has little context to work from, so the range of plausible rewrites is wider.
None of this variance is enough to change a detection outcome reliably in one direction. If it were, that would be a bypass claim, and we do not make one.
If you need a specific result to stay stable — a passage you have already approved — copy it out rather than re-running it. There is no locking mechanism, and the safest way to keep a version you like is to save it yourself before running anything else.