Auditing CGMacros: Interpolated Glucose Series Leak the Future, and a Leak-Free Participant-Held-Out Benchmark for Glucose Forecasting

Public datasets let personalized-nutrition models be compared only if their preprocessing is documented and free of leakage. We audit CGMacros, a free-living dataset of continuous glucose monitoring (CGM), meal photographs, activity and laboratory markers for 45 adults. The dataset paper states that the 1-minute Libre Pro series was linearly interpolated from 15-minute readings; we verify this in the released files (93.7-96.2% of second differences are exactly zero in every participant) and show what it implies for forecasting: any feature computed at an intermediate minute contains the next reading, which we recover to within 10^-6 mg/dL from two consecutive minutes. In a controlled experiment that changes only the CGM features, using the interpolated series raises the pooled R2 of CatBoost from 0.49, 0.38 and 0.35 to 0.86, 0.63 and 0.42 with the same features and 0.89, 0.71 and 0.46 with minute-level features at +15, +30 and +60 minutes, and gives no benefit at 3 and 6 hours. We release a leak-free, participant-held-out benchmark with native-reading extraction, fold assignments, baselines, participant-level inference, spike endpoints, calibration and subgroup analyses. On it, CatBoost with all features reaches R2 of 0.35, 0.47 and 0.41, but a CGM-only linear model already explains 0.23-0.34, and activity, clinical and image features add at most 0.03. Code and documentation are released; the data are not redistributed.