Empirical GNSS Clock-Comparison Correction
Results and validation (no method disclosure)
Summary
We report an empirical correction for long-baseline GNSS clock comparison with typical held-out RMS reduction around ~45%, and strict-window runs up to ~81%. Post-correction residuals are commonly in the 17-22 ns range. The effect has been observed across three independent analysis centers (IGS, CODE, ESA) and multiple years (2023-2025).
This document does not disclose the form, parameters, or implementation of the correction. It describes only the results and the validation protocol, suitable for blind pilot verification by an independent lab.
Evaluation frame (definitions only)
These definitions specify what is being measured and compared. They do not describe how the correction is computed.
- What is transformed (interface level): For each run, we work on an agreed primary comparison or residual time series derived from the submitted clock-comparison inputs (exact series definition is fixed per configuration at onboarding or in the written validation protocol). Deliverables are post-processed series and/or residuals and summary statistics on that object—not a description of internal state or steps.
- Comparator / baseline: “Uncorrected” / “before” is the same submitted stream and the same metric definition (principally RMS of the primary series unless otherwise agreed) computed before Banlys confidential post-processing. “Corrected” / “after” is that identical metric applied to the outputs we return. The comparison is not against a different dataset, relabeled truth, or a re-tuned model on the evaluation window.
- What “improvement” means: Reported percent reduction is the relative improvement in that agreed primary statistic from before to after post-processing. Where we cite held-out results, the correction is not refitted on the evaluation window—consistent with §3 below. Magnitudes are as observed in the documented windows and products; no claim that every deployment will match a single number.
- Form of results (no mechanics): Deliverables are evaluation outputs—time-aligned post-processed series and/or residuals in agreed formats; summary statistics on those outputs (principally RMS, plus any metrics named in the protocol or agreement); consistency- or deviation-style indicators where specified; and optional brief review-level signals for internal review. These are not described as formal calibrated uncertainty products unless explicitly stated in a written agreement. Nothing here specifies how results are produced.
1. Domain and Data
- Domain: Long-baseline GNSS clock comparison (time transfer between distant clocks using final clock products).
- Data: Multi-baseline, multi-day comparison means over multiple years (2023–2025).
- Scope: Effect appears across multiple analysis centers (ACs); not dependent on any single center.
2. Main Result
Typical held-out RMS reduction ~45%; strict-window runs up to ~81%; post-correction residuals commonly 17-22 ns.
- Uncorrected RMS: Typically in the low 30 ns to ~116 ns range on tested datasets, depending on window and baselines.
- Corrected RMS (held-out data): Typical reduction is ~45%, with strict-window runs up to ~81%; residuals commonly fall in the 17-22 ns range.
- Blind validation principle: All improvements are observed on data never used in fitting; correction fitted only on training data, applied independently on test data.
30-day technical audit (IGS Final, strict holdout)
The following result is from a single 30-day test window. The correction was fitted on data from earlier in 2024 and applied to this test window without refitting.
| Metric | Uncorrected | Corrected | Improvement |
|---|---|---|---|
| Test window | 2024-250 to 2024-279 | 2024-250 to 2024-279 | — |
| Baselines | Long-baseline GNSS | Long-baseline GNSS | — |
| RMS residual | 115.6 ns | 22.1 ns | 80.8% (strict-window run) |
| Validation | — | Strict holdout | Verified |
Audit note: This demonstrates high temporal stability: no refit on the 30-day window. Proprietary method. Patent pending (priority Feb 2026).
Note: No claim is made that the residual is zero, or that the correction applies universally across all products. Only observed effect sizes in tested configurations are reported.
3. Validation Protocol
All checks were performed with strict train/test separation; no test data were used in fitting the correction.
| Check | Description | Outcome |
|---|---|---|
| Holdout | Correction fitted on one subset; applied to independent days. | Typical ~45% RMS reduction on test days. |
| Forward-only | Fit on earlier days; test on later days. | Improvement holds; no future-data leakage. |
| Multiple years | Same correction applied across 2023–2025. | Comparable improvement; stable across years. |
| Additional checks | Robustness checks with alternative train/test splits and withheld baselines; applied to unseen data. | Effect is real and generalizes; not driven by a single subset or temporal split. |
A before/after demonstration on public IGS data is available: See demonstration.
These checks confirm the effect is real, reproducible, and generalizable across analysis centers and multiple years.
4. What is Not Disclosed
We do not disclose:
- Functional form of the correction
- Any parameters or coefficients
- Implementation or code
- Anything that could allow reverse-engineering
This ensures trade-secret and patent protection while allowing blind verification.
5. Blind Pilot Proposal
We propose a low-risk, results-only blind pilot:
- The lab provides its own GNSS clock-comparison series in a format we can process.
- We provide a corrected series (or corrected residuals) without revealing the method.
- The lab compares uncorrected vs corrected RMS on their own data to verify the effect.
- Data and results remain confidential under mutual agreement.
- The method is never disclosed; IP and patent status are fully preserved.