Protocol
The questions are asked here before the data exists. It is the only way that a gap, should one come, will mean anything.
Published 10 September 2026 · any question added will carry its date
The protocol in plain words
Before tossing a coin a thousand times, you write on a sheet what you are going to count, and from when you will say that something is happening. You date the sheet, and you never change it again. That is what a protocol written in advance is (fig. 1). Without it, one could always find, after the fact, a figure that suits.
For the colours, the question is simple: over a hundred thousand attempts, is the right colour found more often than chance, which gives one time in four? We also look at whether people do better when they say they are calm.
For the midday point: over ten thousand points, are the points people placed closer to the drawn point than chance allows?
A clarification, added on 16 September: chance does not give the same distance everywhere. From the centre of the square, a drawn point falls on average at 382 out of a thousand; from a corner, at 764 (fig. 2). Each point placed is therefore compared with what chance gives from the exact place where it was placed.
Nothing is published before these numbers. A gap will only count if it is large enough for chance to explain it very poorly: three standard deviations, as statisticians say. No result is published about a person.
The text that follows is the one of 10 September 2026, as it was published. Nothing in it is changed; what is added is dated.
Since 12 October 2026, the Laboratory has been called “The exercises”, The Experience “The colours”, and The Prediction “The midday point”. This protocol keeps the names under which it was published on 10 September 2026; nothing else in it changes.
On 18 October 2026, the minute of preparation was removed from both exercises. The questions about it stop with the attempts made up to 17 October, which remain counted.
The Experience — four colours
What is measured
At each attempt, a colour is drawn at random by the server from four, sealed by a fingerprint published before the choice, then revealed. The person chooses before seeing. A series counts twenty attempts; chance gives five.
What is looked for
Three questions, asked before the data exists. First: does the whole body of attempts exceed chance? Second: do the series made in a declared calm state, the state being declared before the series, exceed those made in a restless state? Third: do the series preceded by a minute of guided preparation exceed those that are not, the preparation being assigned at random?
What will count
No result will be published before one hundred thousand attempts. A gap will be held to be a result if it exceeds three standard deviations, together with the number of attempts and the confidence interval. No individual figure will be published; accounts enter the observations only beyond two hundred attempts. The draw is checked regularly by a uniformity test, published here.
What is not done
No completed series is withdrawn. The questions are not changed along the way; if any are added, they are dated here.
Centre EPÉU, 10 September 2026.
The Prediction — one point a day, added 10 September 2026
What is measured. Each day at twelve twelve, the server places a point at random in a square of a thousand by a thousand, by two independent draws. Before noon, everyone has designated in the same square the place where they felt that point would be. Two measures are kept per prediction: the quarter of the square, right or not, and the distance between the designated point and the point drawn, expressed out of a thousand.
The distance expected at random. Between two points drawn independently and uniformly in a square, the mean distance is two, plus the square root of two, plus five times the inverse hyperbolic sine of one, all divided by fifteen, that is 0.521405 of the side; on this grid, 521 out of a thousand. The value is computed once, published here, and verified by simulation over three hundred thousand draws.
Correction of 16 September 2026 — where chance is measured from. The value of 521 holds between two points drawn at random. The designated point, however, is placed by a person, and its position changes what chance gives: from the centre of the square, a draw lands on average 382 out of a thousand away; from a corner, 764. The comparison retained therefore becomes, for each prediction, the distance expected from the designated point itself — the expectation of the distance between that point and a uniform point of the square, computed in advance and checked by simulation — and, for the group, the mean of those expected distances. That mean gives back 521 when the designated points are themselves uniformly spread. The quarter of the square does not change: it is one chance in four from any point. No data was analysed before this correction, written here before any cumulative figure.
The draw and its seal. The point exists nowhere before twelve twelve: it is drawn at that instant by the server’s task, sealed at once by a thirty-two-byte secret and the SHA-256 fingerprint of the pair, written in a single row, then revealed. Neither the administration nor anyone else can bring it forward, know it beforehand, or change it afterwards; the database itself refuses.
The group result. It appears only from thirty points placed for the same draw. Below that, the page states the number of participants, and nothing else.
What is looked for. The group’s points are closer to the point drawn than chance allows; and more so after a minute of preparation, that preparation being assigned at random one time in two.
What will count. No cumulative figure will be published before ten thousand predictions. A gap will be held to be a result if it exceeds three standard deviations, together with the number of predictions and the confidence interval. No individual figure will be published.
Centre EPÉU, 10 September 2026.
The check on The Experience’s draw, today
This distribution is read on the played attempts of completed series, and is recomputed at each visit. It says nothing about hits: the protocol forbids publishing any before one hundred thousand attempts.