Method
What the tests measure, how the clock runs, and where the curve behind every percentile comes from. Every number on this site is either measured here or an assumption declared as one; this page says which is which.
How the clock runs
Every test runs in your browser and times your input on the spot; nothing is sent away to be scored. A timed test starts its clock on the frame the screen actually changes, tied to the browser's own paint, and reads your press as the key or finger goes down, not as it comes back up. A false start never counts. Most tests average several attempts, because one attempt is mostly luck. The lag of your own screen and input device is part of the number: a 60 Hz panel and a touchscreen read slower than a 144 Hz monitor and a wired mouse, same brain. A press under 100 ms is not a reaction, because nothing human crosses eye to hand that fast; the reaction test counts it as a false start and repeats the run.
The curve behind every percentile
Each result card places your score on a curve and reads a percentile off it. That curve is a histogram with fixed bins, and it begins as an assumption: a starting shape worth a fixed number of assumed runs, taken from published norms or from another site's published histogram (the list below says which, test by test). Every run played on this site adds one run to its bin, so the measured runs outweigh the assumption as they accumulate. A test whose starting shape rests on published data begins with 1000 assumed runs; the four with no published norm behind them begin with 300, so they yield to the measured runs sooner.
The percentile is the share of the pool on the worse side of your score, plus half of those tied with it. One pool per test: phone and desktop, every language, first runs and repeat runs, all in one curve. On the timed tests that means a phone run reads a little slower than the player is and a desktop run a little faster; compare yourself with yourself on the same setup. Only the first completed run per device per test per day is added, so practicing all afternoon counts once. And one network adds at most two runs to any one bar of a test's curve on any one day, so no single connection can pile up a bar: on this page, measured runs means measured runs after that cap.
The starting curve, test by test
For each test: what the starting shape is and where it comes from, how many assumed runs it is worth, and how many runs this site has measured so far. The measured count is read live from the site's own records; it is what turns an assumption into a measurement.
Reaction Time
Log-normal fit, median 273 ms, σ 0.22, an assumption from published norms.
Choice Reaction
Log-normal, median 500 ms, σ 0.22, set around lab four-choice reaction times plus a browser's screen and input lag.
Go / No-Go
Normal, mean 53.4 s, SD 6.7, derived from the test's own timings and an assumed foul rate, with no published norm behind it.
Anticipation
Log-normal, median 55 ms, σ 0.45, an assumption from what lab timers report for adults (a few tens of milliseconds) plus a browser's screen and input lag.
Aim Trainer
Log-normal, median 500 ms per target, σ 0.30, an assumption from what pointing studies report for a mouse at this distance-to-size ratio (Fitts's law) plus a browser's screen and input lag. Mouse and thumb share the one pool.
Stroop
Log-normal, median 750 ms, σ 0.22, set around lab four-choice reaction times plus a browser's screen and input lag plus the Stroop interference.
Typing
Log-normal, median 45 wpm, σ 0.45, an assumption from what typing tests report for everyday typists, the spread widened to match Human Benchmark's typing histogram.
Number Memory
Human Benchmark's published histogram of saved scores, shifted down one digit for self-selection, standing in as the assumed runs (median 8 digits).
Sequence Memory
Human Benchmark's published histogram of saved scores, shifted down one level for self-selection, standing in as the assumed runs (median 8 levels, a long right tail).
Visual Memory
Human Benchmark's published histogram of saved scores, shifted down one level for self-selection and converted to tiles, standing in as the assumed runs (median 12 tiles).
Chimp Test
Human Benchmark's published histogram of saved scores, shifted down one number for self-selection, standing in as the assumed runs (median 9 numbers).
What the curve cannot tell you
Every measured run is taken at its word. The site checks that a posted score is physically possible for that test, a floor and a ceiling for each one, so no bar on a curve rests on a score no human could have produced: a reaction under 100 ms, or a typing speed only a script reaches, is refused and never counted. That removes the impossible, not the dishonest. A believable score that was typed in rather than played looks exactly like a played one, and the site does not trim outliers. What limits the damage is the cap above, at most two runs per bar per network per day, which is why one person with a script cannot take a percentile. The once-a-day rule rests on an identifier the browser keeps for itself, so a cleared browser, or a second one, is a new device to it. Most people play; a few will not, and the curve carries them.
The pool is whoever plays here. People who search for a reaction time test and finish a run are not a sample of everyone: younger, more gamers, more curious, and more repeat visitors as time goes on. A percentile on this site compares you with this site's players, not with people at large. Human Benchmark's saved scores are shifted one level down for that reason; our own runs get no shift, because there is nothing to shift them against.
While the measured count is small, the percentile is mostly the assumption. Against 1000 assumed runs, a hundred measured ones move the curve a little and a thousand share it half and half; the count in each entry above says where a test stands. Until the runs arrive, a percentile is a guess with a number on it.
The clock is as fine as the browser and the device allow. A timed score is read to the browser's timer resolution (a tenth of a millisecond in some browsers, a whole millisecond in others), and your press is sampled by the screen or the mouse at its own rate, a few milliseconds apart on a touchscreen. Below that, the digits are noise. And the number is you today: sleep, caffeine, the hour and practice all move it, and repeat runs sit in the same pool as first ones.
What the site records
Sources
Human Benchmark's published statistics and saved-score histograms (captured 2026-08-27) anchor seven of the eleven starting curves; they measured them, we shifted them, and the shift is an assumption. The other four rest on the reaction-time, pointing (Fitts's law) and typing literature as recalled, plus an allowance for a browser's lag. None of it is a diagnosis or an assessment: the tests are for curiosity and for comparing yourself with yourself.