PINNs, least squares and Bayesian fits all return a parameter. None of them tell you, with a proof, whether a very different parameter fits your data just as well.
Twindex finds those "twins" and proves it. It writes a small JSON certificate, and an exact checker (standard-library Python, rational arithmetic, no floats) replays it: VALID, INVALID or INCONCLUSIVE. A VALID twin pair means no estimator can recover that parameter better than +/- half the gap from those data, under a stated bounded-noise assumption.
Real-data example: the 1978 boarding-school flu outbreak (BMJ). After 5 days of counts, SIR models with R0 = 2.5 and R0 = 20 both fit every count. Twindex identified the next-day count that would separate them, and the real count ruled out both. You can check the proofs in your browser here (runs locally via Pyodide, ~0.2 s per proof): https://arnav-m10.github.io/Twindex/demo/
It also goes the other way: full-prior pavings that prove an upper bound, i.e.…