Accuracy

One number for the whole test set would be the wrong number. Most of the clips are broadcast footage of professionals; the ones that look like what a phone films are the ones that matter, and they are separated out here. Every row is the same build, and the video count is the honest sample size — clips cut from one video move together.

The phone row is the one to read, and the left-hand column is the one that matters: whether it saw you ride away. The right-hand column is the app guessing the trick from the clip with nobody telling it, which is not how Bolts works — it asks what you were going for and measures the go against that. It is kept here because leaving a number out of a page called Accuracy would be the wrong instinct, and because a reader deserves to see what the hard version of the problem costs.

There is a set of 16527 skate clips that people have gone through by hand, marking the trick and the frame each moment happened on.

Bolts never sees most of them while it is being built. Every score below comes from 3128 clips it has never been shown — the same idea as sitting an exam on questions you did not revise from. No figure on this site is typed in; each one is read out of a saved run when the site is built.

The test clips are a mix — 12 filmed at 15 frames a second, 39 at 120 — so the effect of frame rate can be measured rather than asserted.

Counting in frames flatters the slower footage: two frames is a long moment at 15 a second and a short one at 120. Judged in thousandths of a second, which is the like-for-like comparison, the slow-motion footage wins on every moment.

The two greyed columns are the misleading reading; the right-hand pair is the fair one. The peak is the clearest case — it happens between two frames, so a slower camera cannot catch it however good the software is.

Seven targets, written down before the work started so they could not be moved afterwards — 2 met so far, and the rest below with the distance left to run on each.

Open the full run 7 gates, 2 met

Gate	Bar	Measured	Clips	Result
Keeps track of your feet in the air Can it follow your feet while you are off the ground?	≥ 95%	98%	124	clears
Picks the right frame for takeoff Does it agree with a person about when you left the ground?	≥ 85%	80%	124	misses
Picks the right frame for landing Same question for coming back down.	≥ 85%	68%	124	misses
Stays locked on your board Does it follow your board the whole way, instead of wandering onto something else?	≤ 5%	4%	50	clears
Counts the flips right How many times the board rolled over, exactly.	≥ 90%	46%	37	misses
Gets the whole trick right Flip, board spin, your spin and stance — all four at once, or it counts as wrong.	≥ 90%	45%	3128	misses
Knows if you rode away Did you land it, or did you bail?	≥ 90%	80%	3128	misses

Why the whole-trick number is the hard one

Getting the trick right means getting four things right at once — flip, board spin, your spin, stance — and missing any one counts the whole thing wrong. Taken separately the parts score far better than the total. We report the strict version because the generous version would look better and tell you less.

One target we do not score

There is a gate about how well the box around the board lines up. We record the typical overlap rather than the share of frames clearing the line, and the typical value cannot answer that question — so it is absent rather than answered with a number nothing measured.

Test split 3128 clips, 0 errors, setup m7_geom_head, code d1c5008, run 2026-09-15.