A wager fixed before your hand moves

The guess it showed you first

In 1953 Claude Shannon built a relay machine that looked for eight small patterns in a person's play. This page makes the same kind of prediction, seals it, and puts the digest on screen before either button works. Then it can win, miss, and verify the loss in public.

The machine is not reading a mind. It is spending a repeated choice. Use the cryptographic coin and it has nothing to spend.

SHA-256
Current commitment
The published anchor runs before the first wager is accepted.
1. The published anchor, first

The relays have to earn the wager.

Rainer Glaschick built two relay replicas and published a behavioral table for the corrected circuit. The browser runs one cleared machine for 100 moves on each fixed input below. Its random fallback uses one replayable tape, seeded 19530318 from the memo's date. That seed is a disclosed build choice, not a number from the source.

Change ruleFixed inputPublished behaviorObserved, recomputed hereCheck
SLLLL...track after learning94 / 100running
DLRLR...track after learning97 / 100running
SDLLRR...track after learning95 / 100running
SSDLLLRRR...about 2:168 / 100running
SDDLLRLLR...about 2:169 / 100running
SSDDLLLR...at par55 / 100running
SSSDLLLLRRRR...small margin59 / 100running
SDDDLLRLRRLR...small margin58 / 100running

Running 800 fixed moves through the shipped predictor.

Every observed rate is accompanied by its exact 95% Clopper-Pearson interval. “Track” means the final run contains at least 80 consecutive correct predictions, not that random startup guesses were retroactively removed. The two 2:1 rows must exclude 50% while containing two thirds. “At par” must contain 50%. The small-margin rows only have to finish above half.

One discrepancy inside the published source, not with anyone reading it: Glaschick's overview writes the SDD case as “play the same once, then different twice (e.g. llrlllrlll…, rrlrrrlrrr…)”. That example string changes S, D, D, S with period four, so it is not the rule SDD repeated. Repeating SDD produces LLRLLR..., and his own variants report lists this case with the pattern llr.., which is that repetition. The page plays the change rule, and the replica table below shows the same shorter pattern.
How much of this is the tape?

The anchor holds on this tape. It does not hold on every tape.

The eight rows above use one disclosed tape. The button replays all eight predicates on 400 further tapes, xorshift32 seeded 1 through 400, using the same shipped class, and prints the result whether or not it flatters the page. The page never claimed seed independence; this measures how much it would have been overclaiming.

Not run yet. Takes about a second.

The other numbers in the same source

What the two physical relay replicas actually did.

The variants report prints per-replica counts for all nine test cases on its page 8, as machine wins to player wins over runs of different lengths. Glaschick marks a tracked case with an asterisk and prints only the player's wins, so those rows have no rate. The browser turns every counted row into a rate and an exact 95% interval, then asks whether this page's own rate for that case falls inside both. Nothing here is tuned to fit.

CaseInputReplica #1Replica #2This page, same caseInside both intervals
recomputing

Recomputing the replica rates.

A conflict inside the source, stated rather than resolved quietly: the overview says “the independent sequence SSDD (lllrlllrlllr… or rrrlrrrlrrrl… ) is the only one to play at par”, and both physical replicas agree with it (26:26 and 29:23). The schematics-independent software in the same author's variants report puts the same case at 75:25, which is not at par at all. This page lands on the replicas' side of that disagreement, and it lands there because that is what the shipped machine does, not because a side was chosen first.
2. Now it points at you

The hash is on the table.

Choose left or right. The 64-character digest below is computed over a nonce, the round number, the machine's prediction, its active situation, and its reason. Your buttons remain disabled until that digest is visible. After you act, the full preimage appears and is hashed again.

0
sealed round
waiting for the anchor
Your own choices: no rounds yet.
The cryptographic coin: no rounds yet.

Two ledgers, never one. Rounds you choose and rounds the coin chooses are different populations, so they are scored apart and neither rate is ever blended into the other. Both score the same machine carrying the same memory: only the opponent changes. Twenty rounds from one reader are nearly no evidence either way, so the interval stays attached to the score no matter which side is ahead. This visit says nothing about a population.

revealed prediction
not revealed yet
full preimage
not revealed yet
independent shell check
available after the first move
No reveal to verify yet.

Privacy: nothing you choose leaves this browser. This page makes no requests for your moves and stores no session data.

Every round, including every miss
RoundSealChosen bySituationModePredictedChoiceResult
The ledger is empty.
3. The differential control

Change the opponent, not the predictor.

A source with nothing to eat

The button creates a fresh instance of the exact class playing above, then hands that same object to AW.precommit.chanceControl for 2,048 consecutive rounds. Only its opponent changes, to crypto.getRandomValues. Its memory is never reset inside the run.

Not run yet. A valid run has an exact interval containing 50%.

A correct machine still shows a red FAIL THIS RUN on about one honest run in twenty: that is what a 95% interval means, and the red is built in rather than hidden. Press it several times. Chance misses one time in twenty; a machine that could read the generator misses essentially never, which is what the verifier's cheating predictor demonstrates offline.

The way out is on the way in

A real coin does not make the machine lose. It makes both sides approach parity. Use the cryptographic-coin button in the live game and watch that ledger's interval settle around one half. Coin rounds are scored in their own ledger, so the coin's parity is never averaged into the machine's rate against you, and one rate is never reported across two populations. The machine is not failing. It has been denied a repeatable behavior.

prediction needs
same situation + same response + same response

a fresh random bit supplies
no reusable third response
4. The machine can lose

Keep its state, then betray it.

Shannon's memo says the best possible play beats his machine three games to one. The recipe is exact: for each of the eight situations, repeat one same-or-different behavior twice, then change it when the machine is prepared to follow. That requires tracking every cell. This button does so with the same predictor class for 10,000 rounds.

Not run yet.

The seeded random fallback makes this a deterministic replay: the same 10,000 rounds and the same badge every time it is pressed, by anyone. It is a reproduction of a derivation, not a fresh verdict, and pressing it twice is not two pieces of evidence. The target is the player's 75%, and the exact interval must contain it.

Poundstone's useful disagreement

William Poundstone reported doing better on later software games by hiding their score bars and trying to avoid over-switching and short runs. He did not demonstrate a truly random feed, and he did not claim to simulate all eight cells. Glaschick objected that true randomness can only produce parity, while forcing losses needs state tracking. This page separates those claims: the coin control tests parity, and the adversary tests the 3:1 route.

5. What the eight relays remember

Not your side. Your response to a situation.

The machine does not keep a global count of left and right. Before a move it describes the last two results from your point of view and whether you stayed with the same side between them. That makes one of eight situations.

W or Lyour older result
S or Dsame side or different
W or Lyour newer result

Each situation owns two bits of useful memory. The first records whether you stayed or changed the last time this situation occurred. The second records whether that behavior matched the time before. If it matched, the machine predicts you will do it again. If not, the rotating commutator supplied a random choice.

context = (older result, intervening change, newer result)
cell = (last response, was it repeated?)
if repeated: follow it
otherwise: choose randomly

This is a tiny conditional sequence model, not a mind reader. Its leverage is also its weakness: the state depends on the machine's earlier guesses, so a player who follows the state can arrange the history that fools it.

6. What “showed you first” meant in 1953

The chronology has a wrinkle.

Both primary sources give the same order, and both give it plainly. Hagelbarger's page 1: the machine plays first so the player knows it is not cheating, the player announces plus or minus, “(The machine cannot hear.)”, then pushes the play button and the machine lights its lamp, and only then does the player enter the choice on the key. Shannon's memorandum: the player guesses “right” or “left” out loud, the center button is pressed, the machine lights its light, and the player then moves the key switch to the choice already made. So the machine exposed its choice before accepting the electrical scoring input, but after the person had announced a choice out loud. The whole 1953 guarantee rests on a machine that cannot hear and a person's word. This page replaces both with a cryptographic commitment you can check in your own terminal, published before you choose and revealed after.

Shannon's memorandum

A four-page Bell Laboratories internal memorandum, not a journal paper. It defines the eight situations, the two-part memory cell, the random fallback, and the 3:1 best response.

Hagelbarger's SEER

A peer-reviewed IRE article. SEER recorded 5,218 wins in 9,795 plays against Bell Labs visitors and employees. Hagelbarger explicitly warned that short runs, self-selection, deliberate easy patterns, and deliberate cheating biased that ledger. The exact 95% binomial interval is computing, but dependence and selection make it descriptive, not a population interval.

Glaschick's replicas

A self-published technical report by the builder of two relay replicas. He found two contact-label typos and a missing-contact problem; half the memory stayed unused until the third correction. His corrected machines supplied the anchor above.

No population claim: Bell Labs visitors and employees in the 1950s were not a random sample, 9,795 dependent plays are not 9,795 independent people, and your few rounds are one session in 2026. The page reports each ledger for what it is.

The check

What can fail here.

PUBLISHED ANCHORThe page recomputes all eight 100-round rows with the shipped class. Expected hits are 94, 97, 95, 68, 69, 55, 59, and 58. The offline verifier derives them again from the primary-source state rule.
CHANCE CONTROLA fresh, unmodified ShannonMachine carries its memory through 2,048 cryptographic opponent moves. A run fails visibly if its exact interval excludes 50%.
NEGATIVE CONTROLSThe verifier freezes the machine's memory and requires the anchor check to fail. It also gives the chance harness a predictor that can see the deterministic test RNG and requires the interval to exclude 50%.
COMMITMENTEvery choice waits for a SHA-256 digest. The reveal is the exact canonical preimage from the shared kit, verified again in-page and printed with a POSIX shell command.
FREE CHOICESThe anchor uses 100 rounds and a xorshift32 tape seeded from 1953-03-18. The adversary uses 10,000 rounds. These lengths and seeds make the demonstrations replayable; they are not source measurements. The tape matters: the sweep button above replays the eight predicates on 400 other tapes and reports live how often all eight survive.
TWO POPULATIONSRounds you choose and rounds the cryptographic coin chooses are scored in separate ledgers with separate intervals. A single blended rate across both would let the coin's parity be reported as evidence about you, or your rounds be reported as a machine that beats a CSPRNG.
UNCERTAINTYEvery rate in an instrument carries an exact 95% Clopper-Pearson interval. An honest chance control lands outside its 95% interval about one run in twenty, so an occasional red is expected and is not an indictment. Exact binomial p-values assume independent trials, which adaptive rounds are not. They are shown as diagnostics, not repaired into population evidence.
WEB SURVEY, 2026-08-13Four often-cited emulator projects were checked, and every URL tried is printed below with what each returned. Three projects are reachable and one is dead. None shows the machine's guess by default: Williams's page hides it behind the “s” key, the Success Equation remake commits its guess before you press but keeps it in a panel that ships closed, and the AWK program reports its guess after reading your play. Glaschick's own wording is “Some software implementations require to trust the code that only reveals win or loose, not the machine's choice.” Four links measure four links, not how many.
SOURCE LIMITShannon, Hagelbarger, Glaschick, and Breazu were read in full. White's 1959 full text could not be retrieved from its open Elsevier record and is not used for any factual claim beyond bibliographic identification.
Primary record and audit trail

Read what the machine says it is.

Claude E. Shannon, “A Mind-Reading (?) Machine,” Bell Laboratories memorandum, 18 March 1953, reprinted pp. 688-690The interface order, eight situations, two-part cell, random commutator, and best play winning 3:1.
D. W. Hagelbarger, “SEER, A SEquence Extrapolating Robot,” IRE Transactions on Electronic Computers, March 1956, pp. 1-7The play-first interface, 5,218 of 9,795 human-play ledger and its biases, plus the umpired machine duel reported about 55-45 for Shannon over several thousand games.
Rainer Glaschick, “Shannon's Mind-Reading Machine,” 26 August 2024The replica history, correction summary, behavioral specification, and software-interface criticism.
Rainer Glaschick, “Correcting The Schematics Of Shannon's Mind-Reading Machine,” 29 March 2024The contact corrections, core algorithm, simulator variants, and the page-8 measured results from both physical replicas, which are recomputed row by row in the anchor section above.
M. Breazu, D. Volovici, D. I. Morariu, and R. G. Crețulescu, 2020A software implementation of both machines. Across ten 100-play simulations, Shannon averaged 55.8 points to SEER's 44.2; across 200 plays it averaged 112.8 to 87.2. These are simulated machine duels, not human evidence.
Gerald M. White, “Penny Matching Machines,” Information and Control 2 (1959), 349-363The publisher record and abstract were reachable; the full text was not. No technical result from it is used here.
William Poundstone, “How I Beat the Mind-Reading Machine,” 14 July 2014The first-person account of hiding feedback and deliberately allowing longer runs. It is presented as an account, not an experiment with verified random input.

The emulator survey, with its URLs

Four links, checked on 2026-08-13. The question is narrow: does the emulator show you its guess before you commit to yours? Every URL is printed so the survey can be repeated or contradicted.

cs.williams.edu/~bailey/applets/MindReader/HTTP 200. The page is a wrapper for a legacy Java <applet> that modern browsers no longer run, so the game itself was not playable. Its own text: the prediction appears only “If you are interested in seeing the mind reader's prediction of your next press, type 's' (for show).” Hidden by default.
literateprograms.org/mind_reading_machine__awk_.htmlHTTP 200, complete AWK source on the page. Its chunk order is consult model, get input, report results: the guess is computed before the play and printed after it. Inspectable, but not shown first.
success-equation.com/mind-reader/HTTP 200, and this is the live game, one hop from the site's own navigation. It is the companion to Poundstone's subject matter and its copy says the machine “has already guessed what you will choose” and that the guess is “committed before you press.” The panel holding that guess is <section class="hatch" id="hatch" hidden>: a commitment that exists but ships closed. Its predictor is an n-gram vote over your last 0 to 7 bits, not Shannon's eight cells.
success-equation.com/mind_reader.htmlHTTP 200 in about a tenth of a second, but the body served is the site home page, not the game: a soft 404. An earlier draft of this page recorded a timeout here. That was wrong, the correction is this row, and the real game is the row above.
web.media.mit.edu/~guysatat/MindReader/index.htmlHTTP 404. Dead, and not counted as reachable.

Counted: five URLs tried for four distinct projects, three reachable pages, two whose prediction machinery can be inspected, and zero that show the machine's guess by default. That is consistent with Glaschick's “Some software implementations require to trust the code”, and it is far too small a sample to say how many.

Ten Millionths of a SecondThe ancestor that fixed predictions before measurement and kept the ones that died. Always Bet SecondA counter named before the coin falls, but one that cannot lose once the sequences are fixed. The Coin You Can't FakeThe same human difficulty with randomness, detected after the sequence instead of wagered on before it. You Already Know the RestShannon's other prediction game, with a person predicting text rather than a machine predicting a person.