An eccentric 647.6-day model lowers mean squared error in 9 of 18 omitted HD 20794 seasons in a fit with fixed periods and independent Gaussian noise

We test the shape of the known 647.6-day radial velocity signal of HD 20794. We use public HARPS and ESPRESSO data from Nari et al. (2025). Our stated cuts retain 807 nights, one more than the paper. We keep the two inner signals circular. We fix all three periods and fit separate instrument offsets and independent Gaussian noise. We omit each of 18 observing seasons in turn. The eccentric outer model lowers mean squared error in 9 seasons. The mean gain across equally weighted seasons is 0.026 (m/s)^2. A descriptive 95% interval from season resampling is -0.025 to 0.078. After we remove the mean error within each instrument and omitted season, the mean gain is 0.008 (m/s)^2, with an interval of -0.029 to 0.048. In fits to all data, free instrument-season offsets change the best outer eccentricity from 0.507 to 0.111. They reduce twice the log likelihood gain over a circular outer orbit from 19.30 to 1.52. Its eccentricity profile is broad. For that model, twice the log likelihood loss at e = 0.8 is only 0.73. A noiseless projection check shows that these offsets also remove orbital shape information. Thus the weaker eccentric fit gain does not prove a circular orbit or a calibration error. The tests are conditional on the released data processing, chosen periods, and noise model. They are not blind forecasts or a new planet detection. We preserve the source tables, hashes, code, fitted models, scores, and figures.

HD 20794 · radial velocity · eccentricity · season validation · HARPS · ESPRESSO · calibration