A 41.78-day term raises the total prediction score for omitted years in 204 GJ 3998 velocities under fixed noise

We test a fixed 41.78-day velocity term in 204 public HARPS-N observations of GJ 3998. Each of seven year bins is omitted once. All mean coefficients are fitted to the other bins. A fixed correlated-noise model retains coefficient uncertainty and scores each omitted vector jointly. Adding the term raises the summed log prediction score by 35.47 nats. Six bins improve. The 2015 bin supplies about 61 percent of the gain. Prediction RMSE falls from 3.09 to 2.62 m/s. All 13 stated noise and trend cases at 41.78 days give positive totals, from 10.09 to 77.37 nats. The CCF velocities give 18.29 nats under the same LBL noise settings. A simultaneous fit to paired CCF-minus-LBL velocities gives a 41.78-day amplitude of 0.064 m/s. Its nominal 95 percent upper bound is 0.65 m/s with 30-day block resampling. A separate matrix calculation verifies the primary predictions. The package includes the unchanged source table, provenance, code, results, and figures. Periods and primary noise settings come from the prior full-data study. This retrospective test is conditional on those choices. Its score is not a Bayes factor or a false-alarm probability. Shared extraction signals can come from a planet, stellar activity, or instrument effects. The result does not confirm a planet.

GJ 3998 · radial velocity · HARPS-N · prediction · cross-validation · paired measurements · stellar activity