How every course is measured

Everything starts from the line the race actually follows. Where the organiser publishes a course file we use it; otherwise a published GPX of a recent edition, checked against the organiser's map and list of streets. Each source note says which, and for which year.

The original file is kept untouched, with a SHA-256 hash, so a later check is a straight comparison. Before each season every source is fetched again, and any course whose file has changed is measured again.

A line with its points more than 150 m apart on average cuts corners, so it is snapped to the street network before it is measured. Where a course has changed since the file we hold, the page either says so or waits for a new file.

The line is resampled every 25 m, a height is read at each point, and the profile is smoothed over 100 m. The total climb is the sum of every rise along that smoothed profile. Change the sampling or the smoothing and the total changes with it, which is most of the reason published figures disagree.

Each page also shows the raw-to-smoothed ratio: the climb before smoothing divided by the climb after. Around 1.3 means the model is reading the ground. Above about 2.0 the total carries real noise, and the page leans on the net change and the high and low points instead.

A terrain model reads what's under a bridge, usually water, so a river crossing appears as a dive and a climb the runners never make. Through a tunnel it's the other way round: the model reads the street above. Every span where the course runs on a bridge or through a tunnel is found from OpenStreetMap, and where the model has clearly misread it, the span is drawn as a straight line between its ends. The page says which spans were corrected.

What a course costs you against a flat 42.195 km comes from the energy cost of running on each gradient, using Minetti and colleagues' curve (Journal of Applied Physiology, 2002), summed along the profile. A course can climb a lot and still be quick if it gives the height back gently.

It doesn't model the damage a long descent does to your quads, which is why Boston runs harder than its figure suggests.

The weather chart is what race day has been like, year after year. Each race's date rule is applied to every year back to 1995, and the ERA5 reanalysis gives the conditions at the start across the hours of the race. It's the weather on the date the race is held now, so it includes years before the race began and years it didn't run; the source note says where a race has moved.

Records, dates, fees, fields and race-day times come from the organiser, World Athletics and the official results, and each carries the date it was last checked. Before a page is published, a separate check goes through every sentence against those sources and recomputes every figure from the data. Anything that can't be confirmed is hedged or left out.

"The flattest" or "the lowest" is checked across every course measured before it goes on a page, and checked again when a course is added.

A page is due a check the day after its race and again when its entry opens. If you find something wrong, tell us. Corrections are made on the page and the date moves with them.

Every measured course has a GPX download: our resampled line with heights from the terrain model in place of the GPS noise a watch file carries. Use the course lines and profiles however you like. The terrain models and weather data carry their own licences, listed on the credits page.