READ `verdict` AND `evidence`, NOT `match_score`. `verdict` ("match" /
"possible match" / "inconclusive" / "no match") comes from decoy-calibrated
evidence: `evidence.z` (how much better the structure's own predicted shifts
fit the peaks than the same shifts displaced at random), the share of the
spectrum it leaves unexplained, and thresholds set on decoy structures
(`evidence.verdict_calibration`: thresholds, `decoy_exceedance`,
`verdict_reason`). `match_score` is coverage (`match_score_basis`), kept for
existing readers: it rose with the number of peaks and called 31-41% of
constitutional isomers a match on busy 13C lists and every alkane a match on
a real 1H spectrum. A structure with fewer predicted signals than any
calibrated one (about 10 carbons / 6 proton signals) gets "inconclusive";
19F and 31P have no error model and are always "inconclusive".

Omit `tolerance_ppm` and the assignment window follows the nucleus:
1H 0.3 ppm, 13C 3.0, 19F 8.0, 31P 5.0. The evidence does not use it.