The Trusted Referee — Speaker Notes

The pitch in one line: AI has made weather models a commodity — GraphCast, GenCast, Pangu-Weather, and Aurora are all open-weight and runnable by anyone — which makes independent, trusted verification the next scarce resource. RAL, through MET/METplus and the BEACON AI Testbed, is the one neutral institution already positioned to own that role. This deck makes the case to leadership and funders in about fifteen minutes: the shift, the problem, the insight, why RAL specifically, what we'd build, who pays, why timing matters, and a small, concrete ask. Close by asking directly for the seed investment on slide ten.

GenCast — Google DeepMind's diffusion ensemble — beat ECMWF's own 51-member operational ensemble (ENS) on 97.2% of 1,380 target variable/lead-time combinations. A 10-day GraphCast forecast now runs in about 60 seconds on a single GPU, versus hours on a supercomputer for physics-based NWP. Compute and access are no longer the bottleneck. That is exactly why the competitive question has moved from "who has the best model" to "who can be trusted to say which model is best" — and that second question is the one nobody has institutionalized an answer to yet.

These four gaps are real, documented, and each independently would stall adoption on its own. Together they mean a forecaster, an acquisitions officer, or a regulator today has no reliable way to compare an AI model against a physics model, or one AI model against another, on the metrics that actually matter for operations. The literature is catching up piece by piece — Gneiting, WeatherReal, SAFE, DAMBench are each solving one slice — but nobody has assembled the pieces into an institutional standard. MET/METplus itself has no publicly documented AI-model verification extension yet, even though it is the verification capability already designated for NOAA's Unified Forecast System.

This is the crux of the whole pitch. RAL's opportunity is not to out-build Google DeepMind or NVIDIA on modeling — that's an arms race we would enter years late and structurally under-resourced. The opportunity is the one role a commercial lab cannot structurally fill: neutral referee. That role compounds in value as more open models enter the field, rather than being eroded by them — the more models there are, the more the market needs someone credible to sort them. Note for the room: this also requires an explicit governance posture, since RAL's own MILES-CREDIT model would need to be judged by the same neutral bar as everyone else's.

This slide exists to pre-empt the objection "this sounds like a multi-year cold start." It isn't. MET/METplus is already the de facto verification toolkit inside the Unified Forecast System, jointly developed via the DTC with NOAA and the U.S. Air Force. BEACON — RAL's 2025-26 initiative under the WISP program — is already running head-to-head AI-vs-physics benchmarks with more than 30 partner projects. What's missing is not capability, it's a coordinated, funded push to extend that capability to AI-native metrics and turn it into a standing public service rather than a research testbed.

MET-AI is the concrete product, not just the mission. Each piece maps to something RAL can start building this year on top of existing METplus infrastructure. (a) extends METplus itself. (b) is the public-facing trust surface — think WeatherBench 2's role, but operationally oriented and institution-backed, with a diagnostics dashboard doing real work to make it feel authoritative rather than ignored. (c) is what turns this from a research artifact into a revenue-bearing service — vendors and agencies will pay for a credible stamp of approval the same way they pay for other conformance certifications today. (d) is the data-stewardship layer that also doubles as a citable public good and a funding hook in its own right.

This diversity is the funding thesis, not a footnote. No single sponsor — not even NOAA — controls whether this mission exists, which is exactly the kind of resilience NCAR leadership should want in the current funding climate. It also means the service has a real market on day one rather than a hypothetical one: RAL's international precedent already includes partnerships like Kuwait, Korea, and Saudi Arabia, and BEACON's 30-plus partners already span government, industry, and international programs.

WeatherBench 2 is the cautionary precedent, not the model to copy — it standardized research-grade evaluation and became the default because nobody else moved first, but it was never built for operational certification, DOD/FAA trust, or in-situ ground truth. That gap is still open. It will not stay open. Every month without an operational-grade standard is a month some other actor — a commercial lab, a foreign met service, a well-funded startup — can move to fill it instead, and once adopted, verification standards are extremely sticky. UIFCW26 this July is a live venue where this argument is already being made to NOAA EPIC.

Walk through each column briefly: NOAA EPIC funds exactly this kind of community infrastructure; NSF's CSSI program funds sustainable scientific software; DOD/FAA have procurement budgets tied to certification and safety; commercial certification is a real, precedented revenue model — think Underwriters Laboratories, but for forecasts; international met services fund shared standards bodies routinely. The point for leadership: this is one of the few initiatives that gets stronger, not weaker, if NSF core funding is cut, because it was never designed to depend on it.

This is deliberately a small ask, not a moonshot budget, because the entire strategic argument of this deck is that RAL is already about 70% of the way there. Three to four FTE for eighteen months is enough to convert existing METplus and BEACON assets into a public leaderboard and a first certification pilot; it is not enough to build a model from scratch, which is exactly the point — we are not competing to build models, we are standing up the referee. All figures here are illustrative and open to scoping with leadership and funders. The number that matters most is that it is small enough to fund quickly and prove the model before asking for more.

Close on this line and stop talking. The strategic logic of the entire pitch compresses into it: model access has been commoditized by the open-weight AI wave, so the durable competitive position is not owning a model, it's owning trust in the judgment of models — and RAL, through MET/METplus, BEACON, and its DTC relationships, is the one institution already built to hold that role. Ask directly for the seed investment from slide ten and take questions from here.