Adding soft X-ray direction increases spline prediction error for 50 GX 339-4 time bins

We test whether the direction of soft X-ray change improves hard X-ray prediction for GX 339-4. We use public MAXI 2–4 keV and Swift/BAT 15–50 keV data in MJD 60200–60700. Data rules select 50 five-day target bins. Inner calendar blocks select a band delay, spline size, and ridge penalty. Outer 50-day blocks test predictions. Adding direction increases mean squared error from 6.16 to 9.23 × 10^-4 (count cm^-2 s^-1)^2, or 49.8%. A training-mean baseline gives 3.85 × 10^-4. Direction has more error in nine of ten blocks and in all five scored controls. Three other controls fail the minimum data rule. The package preserves frozen inputs, source hashes, all candidate scores, predictions, figures, and code. Independent spline and least-squares calculations verify the primary predictions. Sparse coverage, soft interpolation across gaps up to 25 days, noisy inputs, and the limited model search restrict the result. This is a retrospective prediction benchmark. It does not establish the absence of physical hysteresis.

GX 339-4 · MAXI · Swift BAT · X-ray binaries · hysteresis · blocked prediction · negative result · reproducible benchmark