Abstract
Abstract
Fréchet ChemNet Distance (FCD) cannot distinguish E/Z stereoisomers, because the ChemNet tokenizer maps the SMILES bond-direction symbols / and \ to the same unknown token; inverting every double-bond configuration in MolT5-large’s ChEBI-20 output (384 molecules altered) changes FCD by 0.001 resampling standard deviations while conditional per-bond accuracy falls from 98.97% to 1.03%, and MOSES and GuacaMol cannot expose the defect because their benchmark data contain no stereochemistry. A stereochemistry-aware correction is monotone under synthetic corruption yet prefers fully inverted real output, illustrating that a metric operating only on unpaired generated and reference embeddings cannot recover per-caption correctness. Across real generators FCD carries no dependable signal about a more-than-3-fold difference in E/Z recall: on the double bonds where MolT5-large and BioT5-base both recover the molecular skeleton, recall is 0.26 versus 0.997, yet FCD does not separate the two models─its ordering is statistically borderline, with a bootstrap interval that crosses zero. The failure is one of uninformativeness rather than of reversal. Where paired conditional references exist, stereochemistry should be reported directly, as coverage, precision, and recall, rather than inferred from a distributional score.