Fixed correlated HARPS noise moves 60-75-day GJ 581 forecast scores toward zero in 54 later velocities

We test whether the disputed GJ 581 velocity signal continues in a later instrument. Four circular terms are trained on all 250 released HARPS rows. All 54 later CARMENES rows are scored without fitting their signal parameters. Orthonormal contrasts remove the unknown instrument offset. The forecast includes uncertainty in coefficients, frequencies, and added HARPS noise. Both hypotheses use the same HARPS model. They differ only in whether one trained term continues. The primary white-noise model gives a log predictive density change of -6.25 nats for the 66.8-day term. Fixed 133-day and 10-day correlated HARPS noise move this score to -1.00 and -2.44 nats with the same test noise. Four training models and six fixed test covariances give 24 comparisons for this term. Most scores are adverse. One weak positive result gives +0.675 nat in a 32000-draw check. A secondary period mode near 71 days has 13 percent of the training posterior but supplies 95 percent of its full forecast density. The three established planet signals improve all 72 control comparisons. This shows that the adverse forecast for the disputed term depends on the noise model. The package contains exact public data, source hashes, code, posterior draws, all scores, integration checks, simulations, figures, and reproduction steps. This is a retrospective conditional forecast test. It does not determine whether GJ 581 d exists or prove an activity origin. Epoch and instrument change together. The fixed noise grid does not cover all stellar noise models.

GJ 581 · radial velocity · predictive density · signal continuation · stellar activity · HARPS · CARMENES · public astronomical data