What AbilityBench is allowed to tell you

A screening questionnaire and a measurement are different objects, and the difference decides what a page may say afterwards. If a page asks you to rate how often you lose your train of thought and adds the ratings up, the number is a summary of your ratings; it moves when your mood moves and cannot be checked against anything you did. If a page times how long you take to cancel a keypress you had already started, the number is an interval you actually produced.

Every page here is built on the second kind wherever it exists. Where it does not — the trigger sounds in the misophonia test are the case on this site — the page says it is a questionnaire, and reports how strong a reaction is apart from what it costs you rather than adding the two into one score.

What this rules out is the sentence that would make several of these pages more popular. No page tells you that you have ADHD, dyslexia, misophonia, a squint or anything else, and none implies it by writing “your results suggest”. A screening exercise run by a website, for somebody who chose to take it because they already suspected the answer, has a base rate nothing like the population one, and a confident label from it would be wrong far more often than the visitor would ever find out.

This is a measurement exercise, not a clinical assessment. It reports what you did on this page against a stated reference and nothing more — it cannot establish any condition, on any page here. Only a qualified professional, working with more than a browser, can make that judgment.

A reference figure printed beside a score here either names an author, a year and a title, or — when it is nobody’s finding, such as a chance level that is true by construction or an organization’s published admission criterion — states the basis you would check it against. There is no third kind. The file the figures live in makes the source a required field with no default, and the helpers that print a figure carry its provenance with it.

The reason for the mechanism rather than a house rule: an invented percentile is the worst defect this domain could ship, because it is the one a visitor believes and repeats. “Studies show”, “research suggests” and “the average person” are not sources; an author, a year and a title are.

The useful half is what happens when a page wants a figure and none can be printed. The gap is recorded with its reason, and the page says what it does instead. The card sort prints no percentile because the published tables belong to the test’s publisher. The misophonia test prints no prevalence because published estimates depend on where each study drew the line between annoyance and distress. The verbal fluency test shows the published word counts only as orientation, because they were spoken to an examiner and yours were typed.

Which brings us to the thing this site cannot do, stated plainly: no page here prints a percentile, and not one of the reference figures these pages use is a mean-and-standard-deviation distribution. Every candidate for one either belongs to a copyrighted instrument or was collected with a pencil, a stopwatch and an examiner across the table, and a percentile computed against the second kind would be arithmetically correct and factually false. So results report a raw score with a reference range beside it, and where a comparison would need a distribution the page says there is none.

Raven’s actual matrices, the Wechsler subtests, the card-sorting manual’s tables, the projective sentence-completion blanks: all of these are copyrighted, and several are kept out of circulation for a reason that has nothing to do with copyright. An item bank whose contents circulate stops measuring anything, because the next candidate has seen the items. A site that leaked them would be damaging the instrument for the people who have to sit it.

So every item on this site was written or generated for it, in the format of the real thing and never copied from it. The trade is stated where it matters: a score from the Mensa-format practice set is not comparable to a score from a supervised sitting, and the page says so beside the number rather than in a footer. What does transfer is the format and the pace.

Mensa, Raven’s Progressive Matrices and the Wechsler scales are their owners’ marks. They appear here only to name the assessment a page prepares for — or, on the block design page, the battery that page is telling you it is not — never in the site name or the logo, and each of those pages carries an explicit line saying there is no affiliation, no endorsement and no reproduction of the published test.

Reaction time is a real measurement in a browser, and a smaller one than the figures usually printed beside it suggest. A trial here is timed from the frame that painted the stimulus to the timestamp the browser attached to your keypress, both read off the same monotonic clock, so the subtraction is legal. Two popular alternatives are not: a timer callback fires when the browser gets round to it and measures the timer queue rather than the screen, and a frame callback runs before its own frame is painted, so an onset taken there is a whole frame early — about 16 ms on a 60 Hz screen, in the direction that flatters the visitor.

What none of that removes is the display pipeline. The compositor, the cable and the panel each add latency no web API exposes, and one frame is about 16 ms on a 60 Hz screen before you are involved at all. Every page here that prints milliseconds states that floor and leads with a figure the floor cannot touch where one exists: the mental rotation test reports the slope of your response time against angle, because a constant delay lifts the whole line and leaves its steepness alone, and the rhythm test leads with the spread of your taps rather than their offset for the same reason.

Two rules follow from the same place. On a timed trial, a response faster than 150 ms after the stimulus is treated as an anticipation rather than a reaction and discarded, with the rule stated. And a trial during which the tab went to the background is void rather than slow: browsers stop or throttle a hidden tab’s frames and timers, so a run that kept counting would record a stimulus you never saw and a long response to it. Pages drop those trials, say how many and why, and where they can, ask them again.

Every number on this site is worked out by JavaScript in the tab you are reading it in. Your answers, your response times and your score are never uploaded, logged or kept — the pages are static files, and there is no API for a test to call. That is also why the tests keep working after you disconnect from the network.

This site asks your browser for nothing: no microphone, no camera, no location, no permission prompt of any kind. A task that would otherwise need to hear you — the verbal fluency test is the one here — takes typed input instead and says why on the page. A result you want to keep is copied to your own clipboard as plain text: the handful of numbers the page printed and the seed that rebuilds the run.

And no result is held back. There is no registration, no interstitial, no partial score with the rest behind a form and no upsell. Much of this corner of the internet takes people through a long quiz and then asks for money or an email address before showing them their own answer; simply showing the number is the whole of the position here.

This describes what AbilityBench does. It says nothing about what a browser extension, a corporate proxy or the machine itself might see — only that this site adds nothing to that list.

Underneath every test sit a few shared modules that decide the things that must not be decided differently on every page: the clock and the frame loop, Web Audio, the statistics, the trial runner that owns what happens between “start” and “here is a number”, and the file the reference figures live in. A page supplies the stimulus and the copy; it does not get to invent its own definition of a valid trial. Pages are generated ahead of time and served as static files, so what a search engine reads is the same text you are reading, with the test itself as the only part that needs JavaScript.

Sources are printed on the pages that use them, in full. As an example of the format, the stopping-time range the stop signal task quotes comes from Verbruggen et al. (2019), A consensus guide to capturing the ability to inhibit actions and impulsive behaviors in the stop-signal task, eLife.