Forced-choice precognition

The paradigm

Forced-choice precognition tests whether a subject can choose an outcome at rates above chance, where the outcome is generated after the choice is made. You pick between two neutral options; the browser generates the result with crypto.getRandomValues() only after your choice has been server-signed; the trial is recorded with full integrity metadata.

This paradigm anchors V1 for five concrete reasons:

• Binary outcomes reduce to a single tap — a clean fit for the web.
• Time-symmetric structure: the choice precedes the outcome, so the intentional choice is the dependent variable.
• A well-characterized null. With true RNG the chance baseline is exactly 50% — no need to estimate it empirically.
• Scalable. A user can accumulate thousands of trials over time with no special setup.
• Meta-analytically tractable. Match rates combine across users and studies straightforwardly.

Historical context

Forced-choice testing has a long history. J.B. Rhine’s Zener card experiments (1930s) accumulated massive datasets at match rates some studies put around 50–55% against a 20% baseline — though these have methodological issues by modern standards. The Princeton Engineering Anomalies Research (PEAR) lab (1979–2007) reported effects on the order of fractions of a percent above chance across millions of trials. Daryl Bem’s “Feeling the Future” (2011) modernized the stimuli and reported 51–53% in conditions where psi was hypothesized — and helped spark psychology’s broader replication-crisis conversation.

Proponent meta-analyses report a small persistent effect: Bem, Tressoldi, Rabeyron & Duggan (2015) pooled 90 replications and reported Hedges’ g = 0.09; Storm, Tressoldi, Sherwood & Roe (2023) updated the forced-choice base to 141 studies and reported a much smaller g = 0.02. The first wave of preregistered failures (Galak et al. 2012; Ritchie, Wiseman & French 2012) arrived shortly after Bem (2011), and the “proponents replicate, skeptics don’t” experimenter-moderator hypothesis has not survived later proponent-inclusive preregistered designs.

How a trial works

Each trial follows the same invariant shape:

1. You start a session and see two neutral options.
2. You choose. Your choice and its timestamps are committed and HMAC-signed by the server before any outcome exists.
3. Only after the signature returns does the browser generate the outcome via crypto.getRandomValues().
4. The outcome is revealed and the trial recorded.

Server-signing the choice before any random bytes exist is what makes the dataset usable for research. Any flow in which the client could know the outcome before committing the choice would defeat the integrity model — so violations of the timestamp ordering are excluded from analysis, randomness is cryptographic, every trial counts (abandoned sessions included), and no preview of the outcome is ever shown.

What an above-chance result would — and would not — mean

The paradigm operationalizes a measurable claim: that a subject’s choice can correlate with an outcome generated after the choice but before reveal. If match rates above chance are observed, the candidate explanations are retrocausal precognition, subtle methodological flaws (always possible — the integrity layer addresses the known ones), or pure chance (the default explanation given current evidence).

Fathom does not adjudicate among these. It reports the statistics; whether an above-chance result reflects anything beyond chance is a separate, philosophically charged question the paradigm does not resolve.

The state of the field

The state of evidence on forced-choice precognition is contested. Proponent meta-analyses report a small persistent effect; critical reanalyses (Wagenmakers et al. 2011; Rouder, Morey & Province 2011) find that effect size insufficient to overcome appropriate prior skepticism. The largest preregistered multi-lab replications to date — Maier et al. (2020), N = 2,004, BF₀₁ = 4.38; Kekecs et al. (2023), the Transparent Psi Project (an adversarial collaboration with Bem on the consensus panel); and Walleczek et al. (2025), 26,483 participants across 420,472 trials — have not replicated the effect at high power.

Strong claims in either direction — “the reality of psi” or “psi is impossible” — exceed what current evidence supports. The honest position is that careful researchers continue to disagree about whether there is a residual effect on the order of one percentage point above chance after correcting for biases, and that the question is not closed. Fathom inherits the posture of the Hyman & Honorton (1986) Joint Communiqué: name both sides, quote numbers, end agnostic.