← Projects

AURA

I developed it at Ineco · NASA award to the team

My architecture for verifiable operational decisions in critical infrastructure when weather forecasts disagree.

One weather model forecasts strong wind for the afternoon and another calls it a breeze. Whoever has to decide whether an operation in critical infrastructure goes ahead usually keeps one of the two, or the average, and the disagreement, which was the most useful thing in front of them, gets lost along the way.

AURA is the architecture I developed at Ineco so as not to throw that disagreement away. Instead of keeping a single forecast it works with everything that could happen, and every decision comes with a certificate that says how badly it could go.

An average explains nothing

When a decision touches safety, you need to be able to explain afterwards why it was taken and with what margin. An average gives you a number and keeps quiet about everything else.

That is why AURA has three layers. The first gathers different data sources into credal sets, which keep every value compatible with what the sources say, using the evidence theory of Dempster and Shafer. The second propagates that uncertainty through a structural causal model with a declared error bound for each equation. The third certifies each decision against safety specifications.

A certificate with small print

The guarantee comes with conditions, and they are written down on purpose. Under four explicit and falsifiable hypotheses, which the preprint calls H0 to H3, the computed sets contain the true state end to end. Outside those hypotheses the guarantee does not hold. Every decision carries a two-level certificate, a worst-case status and credal probability bounds on the unsafe event.

So that nobody has to take my word for it, the preprint ships a verification artifact with the engine, 546 tests and 102 properties tested at 20,000 cases each. Every quantitative claim in the paper can be regenerated by a named command.

What v3 withdrew

I am the author of the preprint, and its version 3 dates from August 2026. In it I withdrew the deployment results, validation studies, economic figures and external evaluations that v1 described and that had not been carried out. Those figures do not hold. The paper came out shorter and truer.

In October 2025 NASA gave the AURA team its Public Service Group Achievement Award. The award belongs to the team.

Fig. 1A quay in the storm

Raise the storm or move the third station toward the end20 · 23 · 24 m/s · worst case unsafe · probability 0.01 to 0.16

AURA is an architecture I developed at Ineco for operational decisions when weather predictions contradict each other. It fuses heterogeneous sources into credal sets with bounded support using the evidence theory of Dempster and Shafer, propagates that uncertainty through a structural causal model with declared error bounds per equation and certifies every decision against safety specifications. Every decision carries a certificate with two levels, a worst case status and credal probability bounds on the unsafe event. In the drawing the cabinet gathers three anemometers into a band and the card ticks the worst case box when the band holds a wind above the threshold. The port, the stations, the storm, the 25 m/s threshold, the numbers and the rule on the card belong to this drawing and not to the preprint, and the cups turn in slow motion.

Links

License MIT (code inside the verification artifact of the preprint)

Related