Clock evidence: we corrected a systematic error in long-baseline GNSS comparison

Proof only — method and parameters are not disclosed.

Result

On strictly held-out data (days and baselines that were never used to fit the correction), applying our method reduces the RMS of the clock-comparison residual by approximately 45% (forward-only holdout: fit on earlier days, test on later). Uncorrected RMS was in the low 30s (ns); after correction, residual RMS falls to 17–20 ns. We used several baselines and multiple test days on IGS final clock products, and the same protocol on other years — the reduction is similar. The correction generalizes.

Null checks

Wrong sign: Reversing the sign of our correction makes the residual worse on the same data. The effect has the right direction.
Shuffle: Shuffling the internal association used by the correction removes the benefit. The improvement is tied to real structure in the data, not noise.

What we're not sharing

We are not disclosing the form of the correction, the fitting procedure, the parameters, or the code. No information sufficient to reproduce the correction is disclosed here. Independent replication of the result is possible by running your data through our correction (under agreement); independent replication of the method is not, by design.

Why it matters

Long-baseline clock residuals are a known issue for time transfer and time scales. A validated correction that cuts that residual by a large fraction on unseen data is a concrete improvement. We've done it. The proof is the holdout and the null checks; the method stays ours.