On the evening of the twentieth of March, nineteen forty eight, the forecaster on duty at Tinker Air Force Base in Oklahoma had promised gusty winds and nothing worse. Just before ten at night a tornado crossed the base and wrecked more than fifty aircraft. It was the most expensive tornado Oklahoma had ever seen. Nobody had forecast it, and the official inquiry concluded that nobody could have.

Two officers at the base weather station, Major Ernest Fawbush and Captain Robert Miller, did not accept that. They spent the next few days pulling charts from past tornado days, looking for a common pattern. On the morning of the twenty fifth they thought the new charts looked like the old ones, and they told the base commander, General Fred Borum. In the afternoon the general asked whether they believed tornadoes were likely near the base. When they said the probability was high enough to justify a warning, his entire reply was: do it. At three o'clock they issued the first operational tornado forecast in history, valid from four to six. At about six a second tornado hit the base. Miller later put the odds of that at twenty million to one.

That is the founding story of severe convective environment analysis: the idea that you can look at the air around a storm, before the storm, and say what it will do. Seventy eight years on, the same idea runs the Storm Prediction Center, most forecast offices, and a good share of the research literature. My argument today is that this method became very good at one question and is still poor at a second, and that the two are confused constantly, including in the founding story.

The first question: given that a storm forms here, what kind of storm will it be? The second: will a storm form here at all? Call them the conditional question and the unconditional one. Here is the picture.

A balloon rises through the atmosphere and reports temperature, moisture and wind at every height. Think of it as a core sample of the air. From that column you compute a few numbers. One is CAPE, convective available potential energy. Lift a bubble of surface air. If it stays warmer than its surroundings it keeps rising, and CAPE is the size of the hill it gets to roll down. Another is wind shear, how much the wind changes speed and direction as you climb the column. Shear is what makes a storm tilt and rotate instead of raining out on itself. A third is the height of the cloud base, a measure of how moist the air near the ground is.

The conditional question was largely settled around the turn of the century. In nineteen ninety eight Erik Rasmussen and David Blanchard took every evening balloon sounding from the United States in nineteen ninety two, matched each to the storms nearby, and sorted them into ordinary thunderstorms, supercells, and supercells with significant tornadoes. Combining CAPE with low level rotation did the best job of separating supercells from ordinary storms, though with plenty of overlap. And the single best number for picking out the tornadic supercells was neither CAPE nor shear. It was cloud base height. Half the tornadic soundings had cloud bases below eight hundred metres. Half the non tornadic supercells had them above twelve hundred.

Five years later Richard Thompson and colleagues at the Storm Prediction Center repeated the exercise with hourly model analyses instead of balloons, on four hundred and thirteen storms. Deep layer shear separated supercells from ordinary storms with no overlap between the bottom tenth of the supercells and the top tenth of the others. Low level shear and low level moisture separated the tornadic supercells. They folded the best ingredients into two products, the supercell composite parameter and the significant tornado parameter, which now sit on every severe weather desk in the country.

Now notice what every one of those studies did. They started from storms. A proximity sounding is, by definition, a sounding near a storm that happened, so every number in those tables is conditional on a storm existing. Rasmussen and Blanchard said so themselves: a sounding can only tell you whether conditions are generally favourable, and the forecaster must then look for the mesoscale features, the boundaries, that decide where a storm actually appears. Thompson warned against using his parameters as magic numbers.

The unconditional question is a different animal, and the reason is a nineteen ninety six paper that I think is still underrated. Andrew Crook ran a cloud model over a heated, moist boundary layer and asked how much he had to change the low level air to switch a storm on or off. The answer was one degree of temperature, or one gram of water vapour per kilogram of air. Those are the ordinary measurement errors of a summer afternoon. The difference between a clear evening and a severe storm can live below the resolution of the observing network. Crook concluded that initiation has limited predictability with the observations we have. Six years later the IHOP field campaign sent aircraft and radars to the boundaries where storms were expected to form, and most missions came back with no storm. Colliding boundaries do not always fire, and dry lines were not even the leading trigger that summer.

Put the halves together and you get a base rate problem. Reanalysis climatologies count how often the air over a place looks favourable, and favourable hours vastly outnumber storms. A high CAPE, high shear afternoon in Oklahoma in May is usually just a hot afternoon. The parameters are honest about severity and silent about occurrence. That is also why they travel badly. Mateusz Taszarek's ERA5 climatology found the tail of the CAPE distribution in the United States runs to twice the European values. A study of Chinese supercells published last year found that the thermodynamic ingredients of the American tornado parameter, CAPE and cloud base among them, barely discriminate there at all, and refitted the whole thing around the lowest three hundred metres of shear. A threshold is a property of a sample, not of the atmosphere.

Chuck Doswell and David Schultz wrote the sharpest version of this critique in two thousand and six. A diagnostic variable describes the atmosphere now. A forecast parameter is one whose current value predicts the weather later, verified on cases it was not built from. Almost none of the parameters in daily use, they argued, had ever been tested that way. Their reviewers, who included the man who built the significant tornado parameter, pushed back hard, and the exchange is printed with the paper. The reviewer says the parameters mark where ingredients overlap, and nobody expects a signal twelve hours ahead. The authors reply: then say so, and stop treating them as forecasts. Both were right. They were answering different questions.

Now the part I find genuinely new. For twenty years the field argued over which layer of low level rotation matters most for tornadoes: the lowest five hundred metres, or the lowest three kilometres. Field project balloons said deep. Model based climatologies said shallow. In twenty twenty four Michael Coniglio and Richard Thompson, and in twenty twenty five Brice Coffer and colleagues, found the reason, and it was not the atmosphere. It was the storm motion you plug in. Every rotation parameter is computed relative to how the storm moves, and without a storm you estimate that from the wind profile with a formula. Swap the formula for the observed motion and the ranking flips. Deep layers win, and storm relative winds that had shown no skill suddenly show a lot. Tornadic supercells move noticeably slower than the formula says and a little to its right. Non tornadic ones move a little faster and well to its left.

The best discriminator between a tornadic and a non tornadic supercell depends on how the storm actually moves, which you learn only once the storm exists. Coffer's paper says it outright: deviant motion may be a cause of tornado potential or an effect of it, and the question is left open. It adds that once a storm has modified its own surroundings, it is not certain what a correct background environment even is. The most refined conditional tool of all is conditional on the storm twice over.

One more layer, from Daniel Chavas and Daniel Dawson in twenty twenty one. They built two soundings with the same CAPE and the same low level shear. One produced a long lived supercell in the model. The other produced a cell that died within an hour. The difference was how the moisture was distributed with height, which a bulk number integrates away. So what is environment analysis for?

It is a sieve, not a switch. It tells you which days cannot produce a strong tornado. It tells you, if storms form, roughly what they can become. It does not tell you whether they will form, and its sharpest instruments quietly borrow from the storm they are meant to anticipate. The Storm Prediction Center made the split explicit this March, when it added conditional intensity groups to its outlooks: one set of numbers for how likely severe weather is, and a separate set for how bad it will be if it happens. Two questions, finally drawn on the same map as two questions. Which brings me back to Tinker.

In nineteen ninety nine Robert Maddox and Charlie Crisp went back to the nineteen forty eight charts. Miller had always said the two days looked alike. They did not. The twentieth of March was, by modern standards, an unremarkable day. The twenty fifth was a textbook severe pattern, one that Miller himself later listed in his forecasting manual as one of the two most dangerous in the country. Fawbush and Miller had correctly diagnosed a favourable environment. What they could not have known was that a storm would form on it and cross the same square mile of Oklahoma inside a two hour window. Maddox and Crisp call the forecast fortuitously successful, with luck both good and bad. The environment was right. The rest was the unconditional question, and it still is.