How to read this, and what it can and can't show
The stream. Draws come from the browser's cryptographic generator (crypto.getRandomValues), 500 a second, each landing on one of the poles with equal chance. It is seeded from your computer's hardware entropy, then expanded by an algorithm; researchers in this area usually prefer a physical source, which could be swapped in later. Bytes are fetched 16 at a time as they are needed, so each outcome is made a few hundredths of a second before it is shown, never stored up in advance. Keep this tab in front while recording, because browsers slow background tabs.
The field. At 2 poles the circle fills toward cyan (1) or coral (0) by the z-score of the last few seconds; a quarter of the way up is about z = 1. At 3 or more poles each spoke of the radar shows how far that pole is above or below chance (the solid ring is chance, the dashed rings are ±2), and the glowing orb is pulled toward whichever poles the stream favours.
Jitter is the same lean measured over a short window and drawn continuously, so you can see the wobble between seconds. Changes of leader: pure chance changes the leading pole most seconds (about 30 a minute at 2 poles, about 50 at 6).
Finish & analyse scores every focus second toward the goal it had at that moment (so switching goals is fine), checks the no-focus seconds against chance, compares streaks and leader changes with what chance predicts, and saves the result to the history. The seed is a SHA-256 hash of the session's raw draws, your focus marks and its start time. The short seed (its first 8 hex digits as a number) can seed other randomness later, such as visualizer stacks or a Strudel pattern.
The research so far. The Princeton PEAR lab reported a deviation of roughly 1 in 10,000 over many years. A three-lab replication in 2000 did not reproduce it, and a 2006 meta-analysis found a tiny overall effect that publication bias could explain. At that size it would take tens of millions of draws (tens of hours) to show, so single sessions will look like noise either way. Sessions and history are kept in this browser. Finished sessions are also shared with the global experiment unless you switch off Share my sessions. Because one person rarely records tens of hours, sharing is how this gets tested: every shared session joins the global experiment on the Results & leaderboard page. More on the idea, the evidence and a protocol to try.