assignment not drawn
0draws revealed
0ones
100maximum
waiting to start
Automatic revealing has not started.
1. Stop when you would stop
One hundred fair bits. Your finger is the stopping rule.
Press start. The browser assigns you to PEEK or MASKED with one random bit, generates the full trajectory from Web Crypto, and reveals it. PEEK sees the conventional running p-value. MASKED sees the same kind of bits but no running statistic. Either may stop after the first draw. Stopping is the only freedom: the run ends at 100, and the outcome, coding, and analysis family are frozen in the sealed file below.
Checking the historical anchor and sealed analysis.
Stopped
Three accounts on the same bounded run
Conventional exact p at your stop: . First pointwise p below 0.05: .
Betting bankroll against fair mean 0.5: .
95% betting confidence sequence for the mean: .
The difference: the pointwise p-value is valid for a named look, not for choosing a favorable look after watching. The bankroll and confidence sequence pay for continuous monitoring, provided the bits have conditional mean 0.5 and every stake uses only the past.
Nothing leaves until you inspect this
This is the exact request body produced by the shared kit. It contains six integers, no free text, demographic answer, user-agent, or precise interaction trace. The service separately enforces an IP/day rate limit; IP is not one of these research fields, and this page does not establish infrastructure-log retention.
Nothing has been sent.
2. The living arm
A counter that expects to begin at zero.
The primary outcome is whether the exact two-sided binomial p-value at the chosen stop is below 0.05. It is recomputed from stored successes and stop count. The client-supplied first-crossing field is never trusted for that outcome.
Loading. No result has been inferred yet.
count pending
The store has not been read.
| display arm | valid n | p at stop < .05 | rate | simultaneous CS | bankroll vs .05 | mean stop n |
|---|---|---|---|---|---|---|
| MASKED | ||||||
| PEEK |
PEEK minus MASKED risk difference, estimate and conservative simultaneous interval: .
Anytime-valid means joint over time: under the stated stable conditional-mean assumptions, the confidence sequence covers at every display moment simultaneously with probability at least 95%. A naive fixed-n interval would cover only at its named analysis time and would quietly spend error each time this page was refreshed.
3. The historical anchor
A table of probabilities, not a table of people.
The 1969 Table 2 scan remained inaccessible. Its normal-series values below are therefore labelled as a secondary 2024 transcription. Four overlapping counts have an independent primary check: under the null, double the upper-boundary probabilities printed in McPherson and Armitage's 1971 Table 2.
Computed versus transcribed
| K | 1971 upper, printed | computed x2 | 1969 secondary | check |
|---|
checking cells
The secondary transcription
| consecutive looks K | ever-crossing probability | status |
|---|
Scope: independent normal increments, cumulative sums, a two-sided boundary at 1.959963984540054, nominal 5% tests, and a look after every observation through K. The 0.142 at K = 5 is not a universal inflation factor. The looks are dependent, so 1 minus 0.95 to the K is the wrong comparison.
4. The Gaussian reconstruction
Make a finite replay, then keep it finite.
This deterministic browser replay generates 10,000 independent standard-normal trajectories and records whether each has crossed the historical boundary by K. It is a software check with Monte Carlo error, not published data and not proof.
Not run. Seed 1969 and length 10,000 were fixed in the sealed analysis.
| K | crossed of 10,000 | estimate | Monte Carlo SE | secondary target |
|---|
5. The sealed instrument
The analysis cannot rewrite its own promise.
The literal SHA-256 below was placed in this HTML before the collective arm was read. Each load hashes the shipped analysis file and compares it with this literal. Recomputing a fresh hash and merely printing it would seal nothing.
e1a8488a2924566cf9ffc03e9c363e529ec77cc6ed3a011efb97c831f6e27b86
checking the live file
off-page command pending
Frozen choices: exact probability-ordered two-sided binomial p, strict p below 0.05, maximum 100, alpha 0.025 per arm, 401 grid points, valid complete rows, PEEK minus MASKED, and the shared kit's predictable hedged betting process.
The check
What this page cannot turn into certainty.
HISTORICAL SOURCE LIMITThe 1969 table image was not read. K = 1, 2, 3, 4, 5, and 20 remain secondary-only transcriptions. K = 10, 100, 200, and infinity are cross-checked against primary 1971 cells. K = 50 is omitted: the 2024 lecture does not print it, and doubling the primary 1971 cell would be a derivation, not a transcription. K = 1000 is omitted as unverified.
TWO MODELSThe historical stream is unbounded Gaussian data. The living stream is bounded Bernoulli data. The betting construction shown for the latter is not applied to the former.
CLIENT RANDOMIZATIONThe browser assigns PEEK or MASKED and generates bits with Web Crypto. The wire stores no seed, bit sequence, server signature, or commitment. Later audit of a submitted assignment or trajectory is impossible.
FROZEN WIREThe 10,000-value trajectory label may collide and is not a participant identifier. The first-crossing field cannot be verified from aggregates, so the primary outcome ignores it and recomputes p at stop from two stored counts.
SELECTION AND SPAMSite visitors are self-selected. Local browser marking and the endpoint's rate limit deter casual repeats, not coordinated manipulation. No result here describes a population of scientists.
ABANDONMENTOnly a stopped run can send a row, and the frozen wire has no first-look server commitment. A reader who abandons a run mid-stream leaves no trace, so outcome-correlated abandonment can bias the completed-only arm comparison and nothing on this page can measure that bias.
CONFIDENCE-SEQUENCE ASSUMPTIONEach arm's betting sequence requires a stable conditional mean. Time trends, bots, learning, or changing visitor composition can violate that. Anytime validity does not repair biased sampling.
WHAT OPTIONAL MAY MEANYou may stop or continue a frozen valid process. You may not change the outcome, coding, exclusion rule, stake using future data, estimand, or analysis family after looking.
WHAT THE BANKROLL IS NOTA bankroll crossing 40 controls an ever-crossing error at 0.025 under its null assumptions. It is not a 97.5% probability that the null is false and is not generally a Bayes factor.
DISCRETE P-VALUESThe exact-binomial test's one-look error can be below 0.05 because attainable p-values are discrete. The page does not call every single look exactly 5%.
BOUNDED SEARCHNo correction or retraction was located in the publisher records and targeted searches checked on 2026-08-21. That is a dated search result, not proof that none exists.
Sources and audit trail
Every table says what kind of thing it is.
Armitage, McPherson, and Rowe, 1969, Repeated Significance Tests on Accumulating DataPublisher metadata and summary read. Full Table 2 remained access-controlled and is not claimed as read.
McPherson and Armitage, 1971, primary Table 2 and appendixUpper-boundary null-row cells on printed page 17 and the recursive calculation on page 25.
Chris Jennison, 2024 Armitage LectureSecondary transcription of the famous two-sided normal-series table, labelled as secondary throughout this page. The lecture prints no K = 50 row, so none is attributed to it.
Howard, Ramdas, McAuliffe, and Sekhon, 2021The modern joint-over-time confidence-sequence guarantee and always-valid testing discussion.
Waudby-Smith and Ramdas, online 2023, issue citation 2024The bounded-variable predictable betting construction implemented by the shared kit.
The Null WorldWhat a fixed-look exact p-value counts before repeated looks change the question.
Always Bet SecondAnother fair coin process where dependence, not marginal frequency, determines the outcome.
Ten Millionths of a SecondA reader experiment that refuses to print a personal estimate before its stopping rule earns one.
The other living experiments of this wave, honestly sized: several arms are open and still hold zero rows. False consensus asks you to forecast this page's own crowd after one forced choice. Guess two-thirds aims your number at an average that moves as rows arrive. The anchoring RCT randomizes one number prompt before a Houston estimate. Serial position scores a private 15-word recall and sends two integers. Change blindness counts the alternations you need with and without the blank, and names its no-pilot caveat outright. Wason 2-4-6 lets you probe a hidden rule with six declared tests. Calibration intervals scores eight subjective 90% ranges for coverage. Wisdom of crowds compares your two dot-field guesses with an accumulating crowd. The coin you can't fake catches hand-faked randomness in the browser and now pools three summary integers per typed sequence. The focal point plays five Schelling coordination games beside a live reader pool.