Skip to the test
HumanBenchTests

Method

How These Tests Work

Every instrument on HumanBenchTests runs entirely in your browser, shows the formula behind its number, and states what that number includes besides you. This page collects the method, the reference data and — more importantly — the limitations, in one place.

Principles

  • The answer comes first. No sign-up, no email gate, no modal, and no advertisement between you and a result. Every test on this site is usable within a second of the page painting.
  • The arithmetic is shown. Where a page computes something, it expands the formula with your own numbers substituted in. A bare figure you cannot check is worth very little.
  • Models are labeled as models.The typing percentile curve and the reaction time distribution are models, anchored on published figures. They are described as models everywhere they appear rather than presented as a survey of this site’s visitors, because inventing data would be worse than publishing none.
  • Nothing is uploaded.There is no server component to any test here. Keystrokes, clicks, tones and plate answers never leave your device; personal bests live in your browser’s localStorage and nowhere else.
  • Health-adjacent pages screen, they do not diagnose. The color vision and hearing pages say so prominently, explain why, and point at the professional who can actually answer the question.

Method, per instrument

  • Typing — one word is five characters including spaces. Net WPM = (characters ÷ 5 − uncorrected errors) ÷ minutes. Errors are counted on the final buffer, so corrections cost time but not accuracy. The prompt is drawn from the most frequent English words, and the first prompt is generated on the server so the test is usable before JavaScript loads.
  • Reaction time performance.now(), with the start timestamp taken inside a double requestAnimationFrame so it lands after the frame that actually presents the stimulus. Five trials, median reported. Responses under 100 ms are treated as anticipations and discarded.
  • Color vision— plates generated from cone confusion lines in LMS space using the Hunt–Pointer–Estevez fundamentals, with the dichromat simulation of Viénot, Brettel & Mollon (1999) used to verify that each pair collapses under simulation. Per-dot lightness jitter is applied in linear light so the confusion property survives it.
  • Hearing — Web Audio oscillator, per-ear stereo panning, gain ramped rather than stepped so no click is ever produced. Reports the highest frequency detected at a fixed level, never decibels hearing level, because a browser has no way to know the sound pressure at your eardrum.
  • Ruler — calibrated against ISO/IEC 7810 ID-1 (85.60 mm × 53.98 mm). Fractional inches round to the nearest sixteenth and then reduce, with the whole-inch rollover handled properly.
  • Aim— time to target plus the Shannon form of Fitts’s law, ID = log₂(D ÷ W + 1), giving throughput in bits per second so that scores are comparable across target sizes.

Every instrument

What none of these can do

None of them can diagnose a medical condition. None of them can substitute for a calibrated instrument in a controlled environment. And none of them can tell you much from a single run — the useful signal in a benchmark is the trend across several sessions on the same hardware, not one number on one afternoon.

If a result here worries you, the right next step is a professional: an optometrist for vision, an audiologist for hearing, a doctor for anything that has changed suddenly.