The measurement defined, the statistical framework stated, and the limits of what it can establish.
One quantity, and only one: the deviation of a debiased bit stream from chance.
The instrument collects physical noise on your device: the timing jitter between independent hardware clocks. Each sample is folded to a single bit and debiased with an alternating template, so any steady tilt cancels itself. What remains should be a fair coin. Bits group into trials of exactly 200; a trial expects 100 heads with a standard deviation of about 7.07. Nothing else is measured. Every figure, note, and game on this site is an expression of how far this coin has run from its expectation, and you can watch the mechanism work, live, or read its code.
Every number here is a comparison against one opponent: pure chance.
The working assumption, always, is that the coin is fair and deaf: that mind makes no difference at all. Statistics can never prove that an effect exists; it can only say how surprising the record would be if chance were the whole story. That surprise is what a z-score measures: the record's deviation divided by how far luck alone should wander. Luck's wander grows like the square root of the flips, so the same z means the same rarity at any scale. From z follows the odds we print: a one-tailed “1 in N”, the fraction of pure-chance runs that would lean at least this far in the aimed direction. About one sitting in three reaches ±1σ by luck; one in twenty-two reaches +2σ; one in seven hundred and forty reaches +3σ. Rarity is never proof. It is a price tag on the claim “this was just luck”.
Chance wanders and returns. An effect, however small, persists. Time separates them.
A single sitting can never establish anything: chance wears every costume briefly, and the most honest instrument will show you beautiful runs that mean nothing. What separates a real effect from luck is persistence. Luck's lead wanders like √n and keeps returning to zero; a real lean, however tiny per bit, accumulates like δ·n and eventually walks out of luck's funnel forever. This is why the site keeps permanent records and why no single number is ever the story.
Four questions are answerable here, each with computable odds. Whether a record's lean exceeds chance. Whether deviation follows pre-declared intention, which is the only clean test of direction, since a lean measured without a declared aim can say nothing about aims. Whether attended periods differ from rest within the same person, the fairest of all comparisons. And whether a record's variance differs from chance, its weather. What can never be known this way: the meaning of one session, stories fitted after the fact, and anything about a person's worth or their agency in the wider world. The coin measures one narrow channel; it is silent about everything else.
The discipline that keeps the answers honest: trial sizes fixed in advance, a debiasing template that cancels anything steady, diagnostics recorded with every session so degraded streams can be excluded later, milestones set before the data arrive so a running score cannot tempt you into stopping at a flattering moment, and a record that keeps every trial, whichever way it fell. The first principle is that you must not fool yourself.
Hardware randomness is settled engineering. Ours is the browser’s version of it, stated plainly.
The classical chain, refined over fifty years from the PEAR laboratory to the random generators in bank cards, is: a physical noise source, a comparator turning noise into bits, a balancing stage cancelling steady bias, and fixed-size trials. Ours is the same chain built from what a web page can lawfully reach. A device carries several independent oscillators: the display controller’s clock, the CPU’s timebase. They drift against each other with thermal phase noise. Each sample, the instrument spins on the high-resolution clock counting reads until it ticks over; that count carries the jitter. The count is folded to a single bit by parity, XOR-balanced against PEAR’s alternating template, and grouped into 200-bit trials. Deliberately absent: any cryptographic hash. A hash makes every stream look perfectly uniform no matter what enters it, which would scrub exactly the small statistical deviations the instrument exists to record. Balancing here is linear and transparent, as it was at Princeton.
Around the core sit the guards. Frames during scroll or after stalls are dropped, so main-thread load can never masquerade as signal. The hidden-screen channel, paced by the audio graph, runs on probation and must continuously prove fair-coin statistics before a single bit is committed. And every recorded segment carries its own diagnostics: sample rate, probe saturation, lag-1 and lag-2 adjacent-bit agreement, 8-bit block variance, and which pacing generation produced it, so degraded streams can be excluded from analysis later instead of trusted blindly. The code is open.
A laboratory calibrates one shielded machine for months. We calibrate the whole fleet, continuously, in public.
Princeton ran 5.8 million calibration trials on a single characterized device before trusting a result. A fleet of unshielded phones and laptops cannot be calibrated that way, so this site substitutes something a laboratory cannot have: scale and openness. Every trial recorded on calm pages, with no directed exercise running and attended trials subtracted, is treated as continuous calibration, and its aggregate must sit at chance. The figure shows that record live, split by platform and pacing, because instrument bias lives in hardware and browsers and would surface exactly where it lives, while a genuine universal deviation has no reason to respect operating-system boundaries. Beside the mean, each group’s trial variance and serial agreement are checked against the fair coin, since correlation, not bias, is the browser’s likelier failure. Anyone can pull the same numbers from the open endpoint, and the deeper structural analyses, serial spectra, signatures, position effects, run live on the observatory.
What this instrument cannot claim, stated in advance, so that its claims mean something.
Provenance first. A dedicated generator samples a characterized quantum process inside a shielded box. This instrument samples timing jitter through a browser that deliberately coarsens and fuzzes its clocks, so the stream is physical but not certified-quantum, and part of its raw material may be the browser’s own deterministic fuzz. The balancing template cancels anything steady, which cuts both ways: no constant hardware bias can enter the record, and no constant influence on the raw noise could ever be seen, a blind spot this design shares with the Global Consciousness Project. Second, correlation. The trial statistics assume independent bits; correlated bits distort variance before they distort means, which is why the serial diagnostics exist, and why claims stand in a fixed order of strength: trials under randomly assigned aims are most defensible, since no fixed hardware structure can follow a shuffled aim; attended-versus-rest within one person next; a lifetime lean after that; and variance readings last. Third, rate: at roughly 35 bits a second, an effect of the size the literature reports needs years to surface in one person, which is why the record is permanent, aggregated, and never reset.
One accusation deserves its own instrument: that the bits simply follow CPU load. That claim is testable on your own device in six minutes. The load test alternates idle and full-core blocks, checks that the load truly bit by watching the spin count collapse, and then asks whether the lean, the variance ratio, or the serial correlation moved with it, at a three-standard-error threshold fixed in advance. Run it, and download the raw trials it decided from.
The Moon: assigned aims, fixed trials, permanent records. A formal experiment worn as a practice.
The protocol descends directly from the Princeton PEAR laboratory's tripolar design: intention high, intention low, and baseline, pre-stated before the data arrive (Jahn et al., 1997). Here it is worn as the Moon. Each trial assigns you an aim: raise the moon, let it settle, or rest. A round is six trials, two of each, in shuffled order, so no aim is favoured and neither you nor the instrument can drift toward the easy one.
A trial is exactly 512 fresh bits, about fifteen seconds. The moon itself shows the raw walk of the coin: heads lead and it rises, tails lead and it sinks. The trial's score is its deviation aligned with the assigned aim, z = ±(ones − 256)/11.3, the sign flipped for settle trials. Rest trials are controls: recorded, never scored. Trials combine by Stouffer's method, Z = Σz / √k, and the cumulative rarity is printed after every trial, alongside your lifetime record's place on the curve of chance.
The milestone is 48 scored trials, fixed in advance: at that size the experiment has about an 80 percent chance of detecting a moderate personal effect (about a third of a standard deviation per trial) if one exists. Smaller effects need patience in proportion: an effect of the size the PEAR archive reports, about one part in ten thousand per bit, needs hundreds of thousands of trials to surface, which is why every trial posts to the permanent record with its aim, its count, and its raw walk, and why the record never resets. A completed run of 48 is one experiment, not a verdict: whichever way it falls, it joins the record and the record continues.
None of this method is ours. It is a century old, and it has survived its critics by adopting them.
Fixed trials and pre-stated intention come from PEAR's twelve-year program (Jahn et al., 1997). Meta-analysis across laboratories comes from Radin and Nelson (1989), with the publication bias counterargument on the record beside it (Bösch et al., 2006). The statistical practice itself was examined in Statistical Science (Utts, 1991), and the United States government's commissioned assessment concluded the effects meet the standards applied to any other area of science (Utts, 1996). Pre-specification of events comes from the Global Consciousness Project (Nelson & Bancel, 2011). The full argument, with the theory it serves, lives on the science of consciousness page, and the history of the experiments themselves, with their critics, on the experiments page.