A fully shared night calibration model increases formal annual amplitude errors by more than 20% at all phases in 31 of 151 HARPS designs

We tested how dependence within a night changes formal radial velocity precision in public HARPS-RVBank version 02 data. We used 88,594 exposures in 27,250 day bins for 151 target names. Each name had at least 100 bins and a span of five years. We held every exposure error variance fixed. A parameter placed part or all of the added calibration variance into one shared draw per target night. With full sharing, 8,097 night means had error standard deviations more than 20% above independent-error values. The median ratio was 1.472 among 8,975 bins with multiple exposures. For a circular signal at 365.25 days, formal amplitude errors increased by more than 20% at every fixed phase in 31 of 151 designs. The model included instrument-era intercepts and a linear slope. Adding 1 m/s of independent night noise to both models reduced the count above 20% at every phase from 31 to zero. For fixed positive mean weights and independent base errors, full sharing gives the largest night-mean variance allowed by the calibration error sizes. The true covariance is not given by the source columns. These are conditional sensitivity results, not measured error corrections or planet detections. The package contains all source fields used, code, tables, figures, provenance, and offline reproduction instructions.

HARPS · radial velocity · calibration · covariance · night averaging · error propagation · public data · reproducible benchmark