MIT's storm generator invents the disaster your records never caught

The interesting part of what the team calls Extreme Event Aware, or eta-learning, is the filter rather than the generator ^[5]. Producing a rainfall figure larger than any ever measured is trivial. The claimed contribution is a statistical procedure that learns from a region's ordinary daily records and maps, discards the scenarios the data says cannot happen, and keeps what survives: events worse than anything observed that are still admissible ^[3]. Chang describes the target as events "riskier than everything that has happened before and yet are still plausible" ^[18]. That relocates the argument instead of ending it. Under current practice a planner asks whether the archive holds a century-scale disaster to train the simulation on ^[6]. Under this method the question is whether you accept the model's line between unprecedented and impossible.

Sapsis gives planners an arithmetic they can act on. Katrina, he says, is a 30-to-40-year event, and the object of interest is the Katrina that arrives once a century ^[7]. Take those figures at face value and the design case sits between 2.5 and 3.3 times rarer than Katrina ^[8]. The New York illustration is blunter. The most extreme rainfall ever recorded in the city is 200 millimetres, and the method is meant to describe the storm that delivers 300: where it hits, how intense it is, how large an area it covers ^[9]. That is 50 percent above the record ^[10], and it is precisely the number a drainage cross-section or a substation elevation gets argued over.

What MIT's account does not contain is a single accuracy figure ^[11]. There is no reported test of whether eta-learning, trained on a record with its known extremes stripped out, reproduces the extremes that were removed. The paper is open access in Nature Communications ^[13], so that test is available to anyone who wants to run it, and it is the one that decides whether these maps go into a design basis or stay in a seminar room. Sapsis's assertion that no method does this efficiently for rare events ^[12] is the team's own reading of the field, which is the kind of statement a rival group tends to answer with a benchmark.

Until those holdout numbers are on the table, the fair description is a method that has moved the burden of proof rather than discharged it: from the gauge record, which a community either has or does not, to a plausibility assumption, which somebody has to defend.