FILE NO. RAL-2026-VER-01 Strategy Brief · RAL / NSF NCAR

The Trusted Referee

An open verification & benchmarking service for the AI-weather era

NSF NCARResearch Applications LaboratoryStrategy PitchDraft July 2026
01 — The Shift

AI weather forecasting went open-weight.

GraphCast. GenCast. Pangu-Weather. FourCastNet. ECMWF's AIFS. Microsoft's Aurora. Every major AI weather model of the last three years has shipped as open weights — runnable by any lab, any vendor, any government with a GPU. In January 2026, NVIDIA's Earth-2 stack made an open, production-grade model pipeline available to anyone.

GraphCast GenCast Pangu-Weather FourCastNet AIFS Aurora Earth-2

The model is no longer the moat.

NSF NCAR · RAL — The Trusted Referee
02 — The Problem

Everyone ships models. Nobody agrees how to judge them.

  • Deterministic metrics unfairly favor physics models over AI ensembles.[Gneiting et al., "Potential CRPS," 2025]
  • Reanalysis, like ERA5, is a poor stand-in for real observations.[WeatherReal, 2024]
  • Global-average scores hide large regional skill gaps.[SAFE, 2025]
  • No standard benchmark exists for ML data assimilation.[DAMBench, 2025]

Operational adoption is blocked on trust, not accuracy.

NSF NCAR · RAL — The Trusted Referee
03 — The Insight

Verification is a public good only a neutral institution can provide.

A vendor can't credibly grade its own homework. A startup's leaderboard is a marketing asset, not a standard. As models become commodities, trust becomes the scarce resource — and trust isn't manufactured in the same lab that built the model being judged.

The reframe

Not "build a competing model." Be the referee everyone accepts.

NSF NCAR · RAL — The Trusted Referee
04 — The Unfair Advantage

RAL already owns the pieces.

A

MET/METplus

The community verification standard, already embedded across NOAA's Unified Forecast System.

B

BEACON AI Testbed

Already benchmarking AI against physics models — 30+ projects spanning government, industry, and international partners.

C

DTC Relationships

Developmental Testbed Center ties run directly into NOAA and the U.S. Air Force.

D

Verification Science

Decades of institutional expertise in the metrics themselves — not just the software that computes them.

We are not starting from zero. We are 70% there.

NSF NCAR · RAL — The Trusted Referee
05 — What We Build

MET-AI: verification built for data-driven models.

  • a

    AI-native metrics

    Proper scoring rules for ensembles, spatial and neighborhood verification methods, dedicated treatment of extremes, and in-situ observations — not reanalysis — as ground truth.

  • b

    Public, versioned leaderboard

    An open benchmark, versioned like software, that any model — academic, national, or commercial — can submit to.

  • c

    Referee certification

    A qualification path operational adopters can point to: NOAA research-to-operations, DOD, and FAA.

  • d

    Open datasets & pipelines

    Reproducible, citable, and free of vendor lock-in, so results can be checked by anyone.

NSF NCAR · RAL — The Trusted Referee
06 — Who Needs It

The referee has many paying constituencies.

NOAAresearch-to-operations
A research-to-operations gate for AI models entering the Unified Forecast System.
DOD / FAAoperational trust
Trust before operational deployment, where certification failures carry safety and mission risk.
Commercial weather firmsmarket credibility
Independent validation is market credibility they cannot manufacture themselves.
International met servicesshared standard
A shared, neutral standard no single national program has to unilaterally impose.
Insurers & energy traderspriced risk
Forecast quality is priced risk — they need a defensible way to compare providers.
NSF NCAR · RAL — The Trusted Referee
07 — Why Now

Whoever sets the standard becomes the reference.

The field is at its adoption inflection point: AI models are good enough to deploy and everyone knows it, but nobody has agreed on how to compare them for operational use. Standards set at moments like this persist for a decade — the way WeatherBench 2 became the default research benchmark simply by being first, rigorous, and open. The first mover on the operational standard doesn't just get early credit. It gets durable influence over how the entire field measures itself.

This window closes.

NSF NCAR · RAL — The Trusted Referee
08 — Funding

Diversified by design — not dependent on NCAR core.

NOAA EPICR2O vehicle
A community-funding vehicle already built for exactly this kind of verification infrastructure.
NSF (CSSI)cyberinfrastructure
Sustained funding for open benchmarks and pipelines as national software infrastructure.
DOD / FAAmission & procurement
Budgets tied directly to certification and operational trust before fielding a model.
Commercialservice & certification
Real revenue, not hypothetical — BEACON already has 30+ partners in the pipeline.
Internationalshared standard
Partnership funding from met services that need a neutral standard everyone can adopt.

Strategically, this earns external money at the exact moment NSF core funding faces a proposed ~40% cut.

09 — The Ask

Seed the referee.

Fund a small team — illustratively, 3–4 FTE for 18 months — to:

  1. Extend MET/METplus with AI-native verification methods.
  2. Stand up the public leaderboard.
  3. Pilot certification with one NOAA partner and one commercial partner.

Clear. Modest. Catalytic.

Figures illustrative — scoped together with leadership and funders.

NSF NCAR · RAL — The Trusted Referee
Closing Argument

When everyone has the model, the institution everyone trusts to judge them wins.

NSF NCARResearch Applications LaboratoryPrepared for RAL leadership & partners — July 2026