Abstract
<jats:p> Self-driving laboratories couple algorithmic experiment selection with automated execution, yet current evaluations emphasise outcome-centric endpoints—best yield, Pareto-front expansion, experiment count—without quantifying how efficiently the underlying model learns per unit resource. We propose Return on Experiment (RoE), defined as the ratio of information gain to audited resource cost. Information gain is operationalised through domain-averaged reduction in Gaussian-process predictive variance; resource cost spans experimental expenditures (time, reagents, monetary cost) and computational overhead (tokens, compute hours). Using a synthetic benchmark on the Branin–Hoo function with Gaussian observation noise, we demonstrate across 60 independent campaigns (20 seeds × 3 acquisition strategies) that RoE discriminates acquisition strategies masked by outcome-only reporting: expected improvement and uncertainty-sampling exploration achieve 1.23× and 1.19× higher median cumulative RoE than greedy exploitation (Wilcoxon signed rank <jats:italic toggle="yes">p</jats:italic> < 10 <jats:sup>−4</jats:sup> and <jats:italic toggle="yes">p</jats:italic> = 8.3×10 <jats:sup>−3</jats:sup> , respectively; medians of paired per-seed ratios). A calibration diagnostic reveals that all strategies can degrade empirical coverage below the nominal 95% level, with worst-case coverage reaching 0.37 (expected improvement), validating the requirement that uncertainty-based RoE comparisons must be gated by calibration assessment. RoE complements existing outcome metrics by making learning efficiency an explicit, auditable, and reportable quantity for autonomous chemistry evaluation </jats:p>