A Three-Level Noise Floor for Cyclodextrin Binding Constants and What It Permits for Descriptor-Based Prediction

Regression and machine-learning models for cyclodextrin inclusion constants are often reported with cross-validated Q2 near 0.9, yet the reproducibility of the compiled data on which they are trained has not been quantified. This work quantifies it at three levels, each defined by the uncertainty it captures. The within-experiment fitting precision of a single calorimetric study, 0.008 log units, is a lower bound on repeatability; head-to-head comparisons between laboratories and methods give 0.115 log units (95% CI: 0.079–0.210); and 128 replicated host–guest pairs in a verified extraction of the standard thermodynamic compilation give 0.157 log units (95% CI: 0.115–0.220, from a pair-level bootstrap), with a heavy tail in which 7.3% of pairwise comparisons, and an estimated 3–5% of single determinations, deviate by more than 1 log unit. Propagated through the replication structure of a 44-pair modeling set under three error models, this noise implies an out-of-sample R2 ceiling of about 0.70–0.78, and response noise accounts for roughly two-thirds to nine-tenths of the leave-one-out error of the adopted model, an error statistically compatible with a model limited by the data alone. Against this bound, a size descriptor derived from measured limiting diffusion coefficients through the Edward-corrected Stokes–Einstein relation was compared with McGowan volume, each paired with log P. The two are statistically indistinguishable (95% CI of the difference in mean squared error, −0.024 to +0.042) under leave-one-out and leave-one-guest-out validation, and both lose accuracy when whole chemical families are withheld. No packing optimum is detectable, and normalizing guest size by cavity volume degrades performance by aliasing the size term onto the host intercepts. The durable result is the bound, which should accompany any accuracy reported on compiled data.