Turns out there is no single best way to fake high-resolution weather data from a coarse climate model - it depends on what you are willing to sacrifice.
Researchers benchmarked five spatial downscaling methods, the techniques that turn coarse climate-model grids into finer-resolution maps of temperature, wind, and precipitation, against ERA5 reanalysis data. Rather than judge the methods with one error score, they scored each on five criteria: pointwise error, structural similarity, distribution error, spectral error, and gradient error. The results split cleanly. Some methods nailed pointwise accuracy and got local structures in the right place, but smoothed away fine-scale variability. Others preserved far more of that high-frequency detail, at the cost of positioning local features less accurately. Method rankings flipped depending on which metric, and which variable, you looked at.
That matters because downscaled climate data feeds real decisions - flood models, crop forecasts, local heat-wave warnings - that need fine-grained detail climate models do not natively produce. If the 'best' method changes depending on which yardstick you pick, then any paper or product reporting a single error number is quietly choosing a definition of accuracy without telling you.
It is a tidy reminder that benchmark leaderboards are easy to game, even by accident, just by picking which kind of mistake you would rather make.