How this works, and what it cannot do. An automotive stack is four coats inside 100 µm. Each interface reflects a little of the pulse, and those echoes arrive 200–450 fs apart — closer together than the pulse is long, so they overlap and no amount of peak-picking will separate them. What makes it solvable is knowing the material of each coat: with the refractive indices fixed, the only unknowns left are the thicknesses, and the whole waveform can be fitted at once.

Thickness scales with the index you give it. An index 10 % too high returns a total thickness about 9 % too low, and the thinnest coat can be out by 25 %. The library values here are literature figures for typical automotive coatings, not your paint. For absolute numbers, calibrate the indices against a panel you have measured destructively.

Uncertainties are deliberately conservative. They come from profiling the misfit, not from curvature, and in validation they contained the true value more often than a nominal 1σ would — wide rather than flattering. A coat thinner than roughly 5 µm contributes almost nothing at this bandwidth and will come back with an uncertainty as large as itself; that is the honest answer, not a failure.

Validation. Tested against synthetic stacks with known thicknesses: exact recovery on clean data, 0.2–0.7 µm per coat at 1 % noise, 1–2 µm at 3 %. Missing-coat detection was correct in every case tried, including two coats missing at once. Not yet validated against a real painted panel — until it is, treat it as a well-tested model rather than a calibrated instrument.