Auditing Spectral Evidence Cited by Large Language Models in Pyrazolone Glycoside Assignment

Abstract Evaluations of large language models (LLMs) in structure elucidation score the answers returned; none of those examined describes checking cited spectral values signal by signal. Every value cited in the justification fields of 72 responses from two LLMs (gpt-5, gemini-3.1-pro-preview), which chose between drawn N- and O-glycoside candidates for two pairs of pyrazolone glycosides from infrared (IR) spectra alone or with NMR spectra, was checked against the labelled values of the supplied spectra and, for the anomeric carbon or proton, against the signal named. Thirteen fields, all with NMR spectra, gave another signal’s value for the anomeric signal or a value without a recorded counterpart, and an aryl carbon near 104 ppm, also printed for the sugar-free aglycone, was given as the anomeric carbon in 9 of 12 committed (non-abstaining) responses with the spectra of 117a, whose panels labelled and marked the anomeric carbon (8 of 12 in an independent second coding; Cohen’s κ = 0.72 over the 36 NMR-added responses, 0.32 over the 22 that contain its value). On the 2-deoxy pair, whose NMR items proved to combine spectra of two compounds, the count was 10 of 18 (19 of 35 overall); a post hoc IR feature, coded once like the field check, was named in 15 of 18 IR-only and 1 of 18 NMR-added responses on the ribose pair, a difference this design cannot attribute to the NMR spectra; and agreement with the thesis assignments, which leave one riboside’s attachment site and the 2-deoxy assignment open, is secondary.