train on one slice · test on another · ship results, not recipes
Before · typical
low 30s
ns · depends on window
After · holdout
17–22
ns · not zero
Typical improvement
~45%
your mileage will vary
Clock comparison you can defend in review.
Long-baseline GNSS clock comparison, evaluated the hard way: we fit on training data, score on holdout, and send you corrected outputs plus summary stats. How we correct stays in-house.
If you already live in IGS finals and 30 s CLK products: we publish the protocol and typical holdout numbers. We do not publish the fit.
What correction looks like
Raw remainder swings wider; corrected output stays tighter. The animation uses synthetic shapes sized to our published typical ranges — not a dump from any one run.
On holdout, the uncorrected line swings (~30 ns class). After Tare, it sits tighter (~17–22 ns class). The curves are drawn to match our published typical ranges — not copied from any one export.
You get corrected outputs and agreed summary metrics — not our fit recipe. For pinned runs, null tests, and the full protocol, see the evidence room.
Need to watch an incoming path against that reference? See TareWatch →
Typical numbers (held out)
public summary · no fit disclosed| Primary metric | RMS on agreed comparison/residual seriesSame metric before and after — apples to apples |
|---|---|
| Uncorrected (typical) | low 30s nsVaries by baseline and window |
| Corrected (typical holdout) | 17–22 nsTested configs; remainder is not zero |
| Improvement (typical) | ~45% RMS reductionRun a pilot on your panel before you promise this |
| Witness | TareWatch → — flags when an independent stream drifts from the Tare reference (not a standalone anti-spoof box) |
| Next step | Evidence room → — pinned runs, null tests, and the full validation write-up |