CEREAL
CEREAL COMPUTER // RECORDS LAB

A forecast for the air between buildings.

· zaigon9

SIGNAL 02 // AI + SCIENCE • SEPTEMBER 16, 2026
CEREAL COMPUTER // FIELD DIAGRAM
01OBSERVATIONS
02MODEL
03CHECK AGAINST REALITY
Cereal’s proposed evaluation loop. Illustrative, not measured forecast data.

What is established

NVIDIA says the work uses CorrDiff and StormCast. The report describes a current model at roughly 2–3 square-kilometre resolution and a goal of greater detail. Do not describe street-level forecasts as an already demonstrated public service.

Source trail: NVIDIA research report • September 15, 2026

The Cereal reading

Our suggested first test is deliberately mundane: pick a location, state a forecast horizon and compare predictions with an independent observation stream. Repeat that comparison over ordinary days and difficult ones. A beautiful pollution map is only the beginning of that conversation.

The opportunity we would explore is a research dashboard that makes uncertainty visible. It could separate model output from observed measurements and show when the two disagree. That is a product concept, not an announced feature of this project.

For a small builder, the first deliverable should be a useful question and a transparent evaluation, not a claim to have solved environmental forecasting. Ask potential users what decision they would change if a forecast were dependable. If there is no clear answer, more resolution may not be the missing ingredient.

EXPLORE means this is worth following. We would want the actual release, its terms of use and evidence from independent evaluation before recommending it for operational decisions.

CEREAL VERDICT // EXPLORE

Morning Transmission 0003. Reporting is attributed above; proposals and interpretations are Cereal editorial analysis.