Assuming the measurements are independent samples from a normal distribution. Which they of course aren't, as measurements of adjacent days are obviously correlated (if they were independent, a 1-in-7000 event could be expected to happen on about 2 days within a 44-year span). Now the question is what the nature of the deviation is.

- How independent are measurements of different years?

- Has there been a systematic change in the distribution mean?

- Has there been a systematic change in the distribution variance?

- Was there a good reason to assume that the temperature distribution would be normally distributed to begin with? (Maybe there are strong non-additive effects.)

In any case, it's clear that assuming the observed temperatures in the 1991-2020 range follow a normal distribution and temperatures outside that date range will follow the same distribution is a bad model of reality.

I'm not sure how you can make that claim with only 29 years of data without making some pretty big assumptions about the underlying distribution.

Within the 1982-2026 span there is an equally negative 3.5 sigmas deviation somewhere (no year labels on the graph). The article doesn't touch on it at all so I have no context as to what it could be. But it definitely suggests 3.5 sigmas is not really 1 in 7000 years.