Limits of geometric packing descriptors in molecular crystal energy ranking

Simple packing descriptors do not consistently reproduce molecular crystal energy rankings across chemically distinct compounds. A retrospective benchmark of 98 deposited crystal structure prediction candidates from seven molecules was assembled from published relative r²SCAN-D3 electronic energies. Forty entries from four earlier systems were combined with 58 entries for indomethacin, nicotinamide and aspirin. Density, heavy-atom proximity, N–H···O/N geometry and planar-ring proximity were evaluated individually and through an equal-weight rank composite, without fitting feature weights. An implementation audit identified incomplete periodic-image enumeration affecting at least one descriptor in 23 structures. Recalculation with geometry-dependent complete image bounds changed 3.5 Å contact counts in 13 structures. On the three additional molecular systems, the corrected four-feature composite gave a macro-averaged Spearman correlation of 0.187, compared with 0.501 for density and −0.004 for 3.5 Å contacts. The composite improved on density for indomethacin but underperformed for nicotinamide and aspirin. Fifty-four of 813 within-molecule pairs had reference energy differences below 0.2 kJ mol⁻¹. This retrospective threshold does not provide calibrated predictive uncertainty or a prospective abstention rule. Metadata-verified energy joins, exact input hashes, representative invariance tests and reproducible numerical outputs accompany the analysis. The results document representation and implementation sensitivities and heterogeneous transfer between molecules. Neither superiority over established energy-ranking approaches nor prediction of experimental thermodynamic stability is demonstrated.