The rules behind the verdict.

An app that evaluates mountain weather risk must be accountable for how it arrives at a verdict. Here is the model running today, how we hold it to account, and where we stop short of claiming skill.

GRAT-K1: a name for accountability

The model running inside GRAT today is GRAT-K1.

The version label is not a marketing badge; it exists for accountability. When we discuss performance or publish measurements, every evaluation must be unmistakably tied to the exact model iteration that produced it. If the engine changes, the version label changes with it.

What the model does

Mountain terrain breaks generic weather forecasts. What appears as a benign shower in the valley can turn into severe convection along an alpine ridge. GRAT-K1 synthesises three distinct layers for your coordinates and elevation:

Official weather warnings from national meteorological services (including DWD, MeteoAlarm and NOAA/NWS), preserved without machine translation.
High-resolution numerical weather prediction models, selecting the finest regional model available for the specific mountain range you stand in.
Surface weather station observations and local barometric pressure trends.

The model resolves these signals into a single verdict for your position — Safe, Watch, or Danger — always accompanied by the primary reason and the underlying data source.

GRAT-K1 is not a nowcast. It weighs how likely storms are where you stand; it cannot track an individual storm cell across a valley or predict that lightning will hit your ridge in twenty minutes. Consumer apps that imply they can are guessing.

The hold-out method

A weather model cannot mark its own exam. To evaluate GRAT without self-deception, our benchmarking follows four strict methodological rules:

Separate selection and exam years
Thresholds are chosen on one set of past storms and scored on a different year we have never opened. We never tune on the data we grade ourselves with.
Pre-registered criteria
Before inspecting held-out outcomes, decision rules and verification gates are written down in design documents and committed to version control. If an idea fails to meet its pre-registered threshold, it is dropped.
Cell-by-cell measurement
Mountains are not interchangeable. Performance is evaluated cell by cell across distinct ranges (such as the High Alps, the Pyrenees, the Taurus Mountains, the Rockies) rather than averaging away regional shortcomings into a flattering global figure.
No claims in unmeasured terrain
Where reliable lightning or ground observations are absent, we state that the area is unmeasured rather than claiming skill. An unmarked warning in an unmeasured region would be dishonest.

What we do not publish

We will publish how well the model scores. We will not publish the recipe: which fields feed it, how they are weighted, where the thresholds sit, or which experiments survived. Most ideas we test fail and are dropped before they reach the app, and naming them would hand the search path to anyone willing to skip the work.

That line is deliberate. A hit rate you can check keeps us accountable; a formula you can copy does not make the mountains any safer.

GRAT-K2 is under evaluation

We are currently measuring a candidate successor, GRAT-K2.

It has not shipped. It is not in the App Store release, and it does not run on your device. If and when GRAT-K2 satisfies its pre-registered verification criteria across all benchmarked cells, it will ship in a separate release under its own name, and its benchmark figures will be published openly on our Accuracy page. Until then, GRAT runs on GRAT-K1.

How to verify our claims

To inspect the ground truth data, satellites, and rules we use to score ourselves, visit Accuracy.