Three séances, 10 minutes each
Run when this page was built, by the same code the scene runs — the cup, the two hands and the lexicon — with a visitor who is touching the cup and not driving it. Nothing on the board moves at all unless somebody is touching it, so this is not a board talking to an empty room; it is the least anybody can do and still be at the board.
Seed 18531213 · 102 marks · 0.17 a second
URWANDALERTSU0DETAIL. THROUGH. EV. RESTQUE 4SUTRBSTUDENTSUNAMIBIASSET. HREFURBISHEDUC ATION, SURSISTSUNORTH. G
Seed 18520417 · 93 marks · 0.15 a second
ET. RESTPSYCHIATRY, FROM, DETBASIC, RFRED, URUGFQUEST, RGRADU3TREAT. SURE, HREFUSRIG HT3REF? THESEQ, GRO
Seed 20120911 · 88 marks · 0.15 a second
RESSENTIAL. GROVERSEAS. THEREFOREST3THEREF ORESTRY, THEREFORESTRY. GRADU1THREADS, SUR GE. USER. H
It will not say any of this to you. The other hand’s wander is seeded and these three seeds are fixed so that the page is the same page twice; the scene takes a new seed every visit and writes it into the address bar, so a séance is linkable while you are in it. What a link brings back is the other hand — the same rhythm, the same impulses, the same starting place — and not the message. The same seed with a hand resting on the cup, with a hand that nudges it, and with nobody touching it produce three different tapes, the last of them empty. That is not a limitation of the link. There is no message anywhere in the code to carry.
What is actually happening
Nothing holds what the board is going to say. There is no queue, no script, no sentence in a variable, and the transcript above did not exist until it was spelled. The scene is three parts and none of them knows a word:
The cup is a rigid body with mass, kinetic friction and stiction, pushed by two spring forces — a fingertip is a spring, not a handle. Stiction is a real branch rather than a large damping term, which is why the first movement is a lurch rather than a drift. The other hand’s force is capped at 1.9, well under what a visitor pressing on it produces, so a visitor who wants to go somewhere else goes there every time. A board that fought back would be a puppet show; a real one is trivially overridden.
Nothing moves unless you are touching the cup, and when you are only resting on it your hand goes slack: a finger on a planchette is not holding a point in space, it is touching an object, so where your hand pulls from slides toward wherever the cup now is. The moment you move, it snaps back to your finger and you are driving — full stiffness, and you win. One rule gives both halves of “press on the cup and rest your hand”.
The other hand wanders. It moves in bursts and then holds still, because that is what a hand does and because a continuously drifting cup crawls along the arc taking every letter it passes. It has no destination. What it has is a lean: it is drawn toward whichever letters near it would plausibly continue what has already been spelled — summed over every mark on the board at once, weighted by nearness and by plausibility. A gradient is not a target. No next letter is chosen anywhere, and which letter it was resolves only when the cup stops.
A mark is taken by dwell, not by contact: the cup has to settle on a mark and stay there. How long depends on how plausible the mark is — 0.28 seconds for the likeliest letter and 2.2 for one no English word wants. Both are finite, and that is the promise the scene makes: park the cup on Q and hold it there and you get a Q, in about two seconds. The board can be slow to agree. It can never refuse.
The ideomotor effect
People at a Ouija board really are moving the planchette and really do not know they are. That is a description rather than a debunking, and it is a hundred and seventy years old.
William Carpenter named it in 1852: muscular movement directed by suggestion, independently of volition. Faraday settled it in 1853 with an apparatus rather than an argument — a table top built in two layers with an index between them, so that if the table pushed the hands the lower layer would move first, and if the hands pushed the table the upper one would. The upper one always moved first. His sitters were not lying. They watched the index and were astonished.
The modern work is stranger than the debunking. Gauchou, Rensink and Fels put blindfolded sitters at a board and asked them factual questions they believed they did not know the answer to. Through the board they were right 65% of the time. Asked to say the answer out loud, the same people were right 50% — chance. The board was reading something out of them that speech could not.
Andersen and colleagues put mobile eye-trackers on twenty pairs at a Ouija convention and found that sitters predict the letter before the planchette reaches it. The result that matters most here is not about either person: tracking the combined eye movements of the two players predicted where the planchette would land about as well as tracking one player who had been told what to spell. The prediction is real, it is as good as deliberate intention — and it lives in the pair rather than in either of them.
They also found that players get better at predicting with each letter spelled: a message begins effectively random and becomes more predictable as the meaningful options narrow. That one is not a metaphor for what this scene does, it is a description of it, so it is in the code. The board’s certainty is a curve rather than a constant — nearly indifferent between letters at the start of a word, committed four or five letters in.
The one mark that looks backward
Every letter is weighted by what could come next. The four punctuation marks are weighted by whether what has already been spelled is finished — a full stop is plausible exactly when the run of letters is itself a word, and unlikely at every other moment. That is the same model asked a different question, and it is what a phone keyboard is doing when it offers you a period.
It does not make the board decide where the words are. The letters still never do, and the tape is still something a reader divides. It makes the board able to guess, at about the rate English guesses, and be wrong.
The lexicon, and why it is not this site
The board’s sense of what is plausible comes from 9,624 English words in frequency order — not a dictionary. That is modelling rather than economising: the words a person is primed for are exactly the frequent ones, and a full 264,000-word dictionary would make ZYGOTE as reachable as THE. Rarity is the signal.
It is plain English rather than this site’s own writing, and that was measured rather than assumed. The site’s prose is statistically ordinary English — its letter distribution matches published frequencies at Spearman 0.989, mean absolute error 0.26 percentage points — so seeding the board with the site would have changed nothing about the output while making the board a ventriloquist for the page it sits on. The board is not a mouthpiece.
Word frequency is damped: weight is rank to the power of −0.55 rather than −1. At the true Zipf slope the top hundred words swamp everything and the board reads like phone autocomplete, which is the one failure mode that would make this feel familiar instead of uncanny. Everybody has a phone; nobody has been haunted.
There is no space bar and nothing decides that a word has ended. The board keeps the longest run of what it has spelled that English can still continue, and drops a character off the front when it runs dry — so HELLO falls back to LO, the next letters continue LOCAL, and the tape reads HELLOCAL. That is not a defect being tolerated. It is what real transcripts look like, and it is why people can read them: the divisions are put in by whoever is reading.
What it measures
The board’s own letters, across two and a half hours of simulated sitting on twelve seeds, correlate with published English letter frequencies at Spearman 0.84, mean absolute error 2.2 percentage points. It is not English and it is not noise. With plausibility switched off — same hands, same seeds, every mark weighted the same — the vowel share falls from 31.6% to 23.9% against English’s 38.1%, and the tape stops containing words. Punctuation lands on about 6% of marks, which is roughly the rate English punctuates at, and it lands where it does because a full stop is plausible exactly when what has been spelled is already a word.
The residual is the board’s geometry showing through. S, T and R come out at about twice their English rate and A and O at about two thirds, because the letters at the middle of the two arcs are nearest to everything and get more traffic — and widening the arcs, which is what makes the board look like a board, costs about two and a half points of vowel share for exactly that reason. Both were chased far enough to be sure: a spring pulling the other hand back toward the middle made the distribution markedly worse at every strength, and an intermediate arc width bought none of the vowels back. Correcting the rest would mean weighting letters by where they sit on the board, which is a thumb on the scale rather than a model of anything.
What nobody has measured
One thing this scene asserts without any external grounding, said plainly because the rest of the page is careful: there is no published account of how a planchette actually moves. No speed distributions, no dwell times, no acceleration profiles. A camera-based tracking system for Ouija research was built and described in 2019, so somebody has the instrument; the kinematics do not appear to have been published.
Which means the cup's feel — the mass, the stiction, how long it holds before it breaks away, how far it coasts — was tuned by measuring this simulation against itself and reading the numbers, not by matching a real one. Everything on this page about prediction is sourced. Nothing on this page about motion is.
Sources
- W. B. Carpenter, "On the influence of Suggestion in Modifying and directing Muscular Movement independently of Volition," Proceedings of the Royal Institution 1 (1852), 147–153 — the paper that named the ideomotor principle.
- Michael Faraday, "Experimental Investigation of Table-Moving," The Athenaeum (2 July 1853) — the apparatus that settled it: a table top in two layers with an index between them. The upper layer, the one under the sitters' hands, always moved first.
- H. L. Gauchou, R. A. Rensink & S. Fels, "Expression of nonconscious knowledge via ideomotor actions," Consciousness and Cognition 21:2 (2012), 976–982 — blindfolded sitters answering factual questions were right 65% of the time through a Ouija board and 50% when asked to say the answer aloud.
- M. Andersen, K. Nielbo, U. Schjoedt, T. Pfeiffer, A. Roepstorff & J. Sørensen, "Predictive minds in Ouija board sessions," Phenomenology and the Cognitive Sciences 18 (2019), 577–588 — mobile eye-tracking of twenty pairs at a Ouija convention. Two findings are load-bearing here: the combined gaze of the two players predicts the planchette about as well as one player who has been told what to spell, and players predict better with each letter already spelled.
- The lexicon is the english.txt resource of npm most-common-words-by-language @3.0.14 — 9,624 words in frequency order after screening — front-coded to 24KB gzipped and shipped with the page. Nothing is fetched at runtime and no service is called.
- No source, and that is the point: there is no published account of how a planchette actually moves — no speed distributions, no dwell times, no acceleration profiles. A camera-based tracking system for Ouija research was described in 2019, so the instrument exists; the kinematics do not appear to have been published. Everything this scene claims about prediction is sourced. Nothing it claims about motion is.