Canonical is not always best. An Evidence Press briefing. Weather forecasts are corrected many times a day by combining a model forecast with new observations. To know how much to trust the forecast, forecasters run a small ensemble of forecasts and measure how they differ. With only a few dozen members, the ensemble suggests links between distant places that are just chance. So forecasters use localisation: they throw away the relationships most likely to be noise. How well that works depends on the variables you do it in. In twenty twenty-two, Snyder and Hakim proposed a special set of variables, called canonical coordinates, in which the exact update splits into independent pieces. They conjectured that localising there gives the smallest possible error. This paper finds that the answer depends on how error is measured. Measured in the ordinary physical variables, as in their own experiments, the conjecture fails. With one observation it fails for every ensemble size, unless the observation happens to line up with a principal direction of the forecast uncertainty: a slightly rotated set of variables always does better. With several observations, it fails whenever the scoring links an observed direction with an unobserved one. Measured relative to the forecast's own uncertainty, canonical coordinates are the best choice for large ensembles. The paper also gives a formula that ranks any choice of variables, and a test of exactly when canonical coordinates win. In a reconstruction of Snyder and Hakim's two-hundred-variable test case, with eight members, canonical coordinates give about fourteen per cent more error than an ideal filter. Plain symmetric whitening cuts that to about six per cent, which is like doubling the ensemble. These results assume the true forecast uncertainty is known and look at a single update; when it has to be estimated from a small sample, much of the gain disappears. The proofs, exact certificates and code are openly archived. This is an anonymous, AI-assisted, unrefereed candidate, not peer reviewed or independently reproduced. Published on the third of November twenty twenty-six, its paper and evidence are linked from the Evidence Press release page. This briefing uses an AI-generated voice and is not additional scientific evidence.