Credibility
Publishable comparisons need traceable evaluation: what was fit, what was held out, what metric moved, and what stayed internal by design.
GNSS clock comparison · holdout discipline · files in, files out
Your clocks disagree. Tare cleans the leftover you can already measure.
Long-baseline GNSS clock comparison still leaves tens of nanoseconds on the table. Tare is a proprietary off-line transformation—calibrated on training windows, scored on held-out days, returning the same files with values updated. Not a method paper. Not a live feed. Not a 90-day experiment.
Published holdout statistics (IGS-class panels, tested configurations): low-30s ns RMS before → 17–22 ns after (~45% reduction). Whether that repeats is what your first cycle measures—you keep before; we send after; you do the arithmetic.
The service: corrected files are the deliverable. Not a certified time source, not an operational live clock, not a performance guarantee on your panel.
ORBIT → AC → EVAL
illustrative
Artwork: constellation → analysis-center clocks → holdout scoring (not live data)
01 · Why this matters
The number is not the hard part. The argument is.
Labs, product teams, and review panels do not fail on having clocks—they fail on being unable to show that the comparison was fair after the fact. A biased window, a leaky train/test split, or an undisclosed fit can invalidate months of work without anyone noticing until review day.
WITHOUT HOLDOUT
WITH TARE
Operations
Wrong timing bleeds into product specs, campaign planning, and legal traceability. Fixing it late is expensive; proving it early is cheaper.
Control
You should not have to reverse-engineer someone else’s method to use post-processed clocks. Tare ships outputs; the transformation stays with Banlys.
Stakes first. Spreadsheets second.
Read how we define fair before you read how far RMS moved.
02 · What Tare is
Proprietary transformation. Files in, files out.
Tare is a holdout-validated post-processing service for multi-AC GNSS clock comparison. You provide agreed inputs (comparison series, windows, product definitions). We apply a proprietary off-line transformation and return the same files with values updated. You keep before. You score after. Movement on agreed metrics is measured under a locked split—not asserted as a universal gain.
| Category | Post-processing service for timing comparison data |
|---|---|
| Transformation | Proprietary; applied off-line; method not disclosed |
| Domain today | Long-baseline GNSS clock comparison (IGS-class products, agreed baselines) |
| Deliverable | Corrected files in the same format. Optional summary statistics if named at onboarding. |
Other timing domains may follow; this service is GNSS clock-comparison only. How a cycle works: Service. Terms: Service terms · FAQ.
Accepted inputs: multi-AC clock-comparison series and agreed IGS-class clock products. Raw receiver logs and streaming feeds are not accepted unless agreed in writing.
STEP 01
You define the panel
AC products, baselines, train/holdout windows, comparison metric.
STEP 02
We calibrate on train only
Transformation is calibrated on train—not on the days you will score.
STEP 03
We score on holdout
RMS (and agreed summaries) on days your team did not use to calibrate.
STEP 04
You receive outputs
Corrected files. Method remains confidential. You score before vs after.
train ∩ test = ∅ · ship outputs, not recipes · typical ≠ universal until your cycle runs
WINDOW DISCIPLINE
train ∩ holdout = ∅
The evaluation contract is the split. If we calibrate on holdout, the number is not yours anymore.
03 · Why you would want this
Less arguing about whose clock is wrong. More signing off a number you can defend.
+
A comparison you can defend
Holdout discipline is the evaluation contract—not a footnote in someone else’s README.
+
Observed RMS movement (panel-dependent)
In published IGS-class holdout runs, reductions on the order of ~45% appear; remainder commonly 17–22 ns—not zero, not guaranteed on your data.
+
No method archaeology
You integrate deliverables; you do not rebuild their entire research stack.
+
Evidence you can open
Pinned DOY ranges, null battery, strict-window examples—for reviewers who read footnotes.
+
A cycle, not a science fair
First paid cycle is the check. If it moves, you continue. If it does not, you stop.
04 · Where corrected time lands
One holdout spine. Different clocks on the line.
Tare is not “an HFT product” or “a telecom product.” It is defensible GNSS clock comparison that many downstream desks, SLAs, and review panels need to trust. The verticals below are who feels the residual first—not separate miracle claims.
Same deliverable today: train/holdout post-processing, corrected files. Maturity tags describe roadmap—not separate performance claims.
Service today
TareHFT · software path
Capital markets & sync desks
Co-lo timing and event ordering live on whose clock you trust.
When residual noise is tens of nanoseconds, sequence calls and latency budgets amplify small disagreements. Tare gives you a holdout-defensible comparison before you bet desk architecture on it. Desk visuals here are illustrative—not field claims.
I/O shape: recorded CLK / comparison panels → corrected series (batch; streaming envelope in TareHFT phase 2).
Service today
Telecom & network sync
5G and backbone sync need traceable time, not heroic holdover stories.
Carriers and timing-as-a-service vendors defend SLAs with comparison data across products and windows. Tare is how you show the panel was scored fairly—before it becomes a field incident.
Same service: AC clock products, agreed baselines, files you can score against the before you kept.
Service today
Academic track
Labs, NMIs & product integrators
Publishable comparisons and shippable clock specs share one bar: fair evaluation.
National labs and GNSS timing vendors need reviewers to see what was fit, what was held out, and what moved—without handing over the correction recipe. Academic and commercial customers use the same job at different list prices.
Evidence room + technical note for footnotes; first cycle for your panel.
Stakes vary. The split does not.
Pick your lane above; see what else we’re building or jump to adoption.
05 · How you adopt it
Hard science. Practical envelope.
Implementation is intentionally boring: files in, evaluated outputs out. No rip-and-replace of your entire stack on day one.
1
Apply / assign links
Tell us your AC products, baselines, and metric. We assign link IDs and lock train/holdout dates before touching data.
2
Secure data handoff
Sneakernet/USB or agreed transfer—your security rules lead. We do not need your method; we need the comparison inputs you already use.
3
Train / holdout run
We fit on train, report on holdout, document nulls and pinned exhibits. You get corrected files.
4
Decide on merit
If holdout holds on your panel, keep sending cycles. If not, you still have a documented run—not a marketing slide.
How you buy it
Fee by assigned link and cycle · academic and commercial list · quoted at onboarding · each link a separate fee · payment before first cycle. Not a live feed. Not a certified time source.
06 · Proof (not promises)
The excitement is real. The numbers are qualified.
Orientation numbers below come from published IGS-class holdout runs—not a forecast for your panel. Pinned runs and null tests in the evidence room are what survive review.
Published vs your cycle: public exhibits use fixed IGS-class panels and the same train/holdout frame as a paid run. They show what we observed on those panels; your first cycle asks whether similar movement appears on yours—representative of the protocol and domain, not a guaranteed effect size.
RMS · BEFORE / AFTER · HOLDOUT WINDOW
animated · illustrative
Illustrative trace — shape only; pinned DOY tables live in Evidence
RMS_before · typical
low 30s
ns · window-dependent
RMS_after · holdout
17–22
ns · not remainder = 0
Δ_RMS · published runs
~45%
median-ish in tested configs · panel-dependent
Pinned strict-window example
80.8%
Labeled reduction on DOY 2024-250–279 (IGS/CODE/ESA context in technical note). One published exhibit—not a claim that every window behaves identically.
EVIDENCE · PINNED_DOYNull battery
holdout ≠ luck
Time-shift and synthetic controls documented in the evidence room. If the effect were trivial leakage, you would see it here.
EVIDENCE · NULLSFootnotes welcome.
Console layout, DOY tables, and scope limits. Full definitions in the technical note.
07 · What we’re building
Multiple Tare programs. One timing bench.
The service is what you can contract today. The family below shares holdout culture and Banlys timing hardware heritage—each line has its own maturity and evidence bar.
North star: a sealed corrector at the timing boundary (live or near-live I/O, method inside the box)—only after holdout proof on your panel and a streaming latency envelope, not before.
Service today
Tare Program
Holdout-validated GNSS clock-comparison post-processing: corrected files, optional RMS pack, evidence room. File in, file out. That is the job.
Software path
Phase 2 queued
TareHFT
Batch validation closed on recorded CLK panels (harness path). Next: streaming envelope, per-epoch latency shape, CGGTTS I/O spec for desk integration.
R&D
TareGEO
Does corrected timing show up in PPP / position? Baseline vs IGS20 runs exist; optional step is correction inside the PPP chain. Separate from core clock-comparison proof tables.
Live desk demo
TareWatch
Witness on a Tare-corrected reference vs independent incoming GNSS stream—alarm when paths diverge long enough (spoof / anomaly). Synthetic demo on site; live two-receiver desk kit validated. Streaming reference at customer sites is a later milestone.
08 · Who runs it
Built by Banlys.
Tare Program is operated by Banlys, LLC—timing hardware, precision tools, and holdout-validated clock-comparison post-processing from the same shop.