# EXP-003 — Forced-choice card guessing with published commitments **Status:** filed, collection not yet open **Protocol version:** v1.0 **Filed:** 2026-07-24 **Experiment slug:** `psychic-self-test` This document is the preregistration. It is published before any data exists, and its SHA-256 is displayed on the experiment page. --- ## 1. Question Can a participant identify a randomly selected Zener card at better than one in five, across a standard run of 25 trials? ## 2. Hypotheses - **H₀ (null):** the probability of a correct identification is exactly p₀ = 1/5 = 0.20. - **H₁:** the probability differs from p₀. Two-sided: a consistently sub-chance rate is also a result and would be reported. It is the classic *psi-missing* claim, and it should not be able to hide inside "no effect". ## 3. The problem this design exists to solve Every online psychic test has the same fatal flaw: the server chooses the card, and the server also decides whether you were right. Nothing stops it choosing the card *after* seeing the guess. A participant who scores well cannot show anyone that the test was fair, and a participant who scores badly can always suspect it was not. The data is worthless in both directions. So this register uses a **commitment scheme**, which is the same idea as sealing a prediction in an envelope before the event. ## 4. Commit–reveal, precisely For each trial, before the participant sees anything: 1. The server draws a card uniformly from the five symbols, and a 32-byte random nonce. 2. It computes `commitment = SHA-256(card || ":" || nonce)` and stores it. 3. It sends the participant **only the commitment**. The participant guesses. Then the server reveals the card and the nonce, and the participant — or anyone reading the published data later — can recompute the hash and confirm it equals the commitment they were given before guessing. **What this proves:** the card was fixed before the guess existed. The server cannot have chosen it afterwards, because it would have had to find a different `(card, nonce)` pair hashing to the same value, which SHA-256 does not permit. **What this does not prove:** that the card was drawn *fairly*. A server could commit honestly to a biased draw. That limitation is stated here rather than glossed: the defence against it is the published data, in which the distribution of dealt cards is checkable — a biased generator shows up as a skewed card frequency across tens of thousands of trials, and §8 makes that a reported quantity rather than an afterthought. The nonce is 32 random bytes so the commitment cannot be brute-forced. With only five possible cards, a short nonce would let a participant simply try every combination and read the answer off the hash before guessing. ## 5. Primary outcome The overall hit rate across all completed trials, tested against p₀ = 0.20. ## 6. Statistical analysis Fixed before collection, not modifiable while the register is open: - **Test:** two-sided exact binomial against p₀ = 0.20, α = .05. - **Interval estimates:** Wilson score intervals at 95%. - **Milestone reads:** at 10,000, 50,000, and 250,000 trials. - **Per-symbol analysis:** five secondary comparisons with Holm–Bonferroni correction, to catch a bias toward one symbol — in participants or in the generator. ## 7. The leaderboard, and why it is a liability A streak leaderboard is the reason people share a test like this, and it is also the fastest way to manufacture a false psychic. So the arithmetic is published beside it. In a single run of 25 trials at p = 0.2: | Score | Probability | Roughly | |---|---|---| | 9 or more | 0.0468 | 1 run in 21 | | 10 or more | 0.0173 | 1 in 58 | | 12 or more | 0.0015 | 1 in 649 | | 15 or more | 0.000014 | 1 in 73,720 | A run of 9 or more out of 25 — which feels remarkable — happens about **once every 21 runs by chance alone**. At 5,000 completed runs, roughly 234 people will score 9 or more without any ability whatsoever. The often-repeated claim that scoring 9 of 10 is "about one in three thousand" is wrong, and wrong in the direction that flatters the scorer: it is one in **238,186**. The one-in-3,125 figure belongs to a different event — five consecutive hits — and the two get conflated. The leaderboard therefore displays, beside every entry, how many people would be expected to reach that score by chance given the number of runs completed. A score is only interesting when it exceeds what the crowd size predicts. ## 8. Generator integrity Dealt-card frequencies are published and monitored. Under a fair generator each symbol appears in a fifth of trials; a persistent deviation indicates a problem with our randomness, not a discovery about participants, and would be reported as such. This check is prespecified so that it cannot later look like an excuse invented to explain away a result. ## 9. Eligibility and deduplication - Unlimited trials per participant: this measures a rate, not a single verdict, and repeated attempts are informative rather than contaminating. - Trials dealt but never answered are recorded and excluded from the rate. They cannot bias it — a participant abandoning a trial has not seen the card — but the count is published so the exclusion is visible. - Participants are anonymous, identified by a device-scoped random identifier. ## 10. Data publication Every trial is published with its commitment, nonce, dealt card, guess, and outcome. That is what makes the fairness claim checkable by a stranger rather than a promise from us: anyone can recompute every hash in the file. ## 11. Prespecified threats to validity 1. **Self-selection.** Participants who score well early are likelier to continue, which inflates trial counts among the lucky without biasing the overall rate. Reported per participant so it can be inspected. 2. **The generator.** Addressed in §8. 3. **Multiplicity across participants.** With enough participants, extreme individual results are guaranteed. No individual result is treated as a finding; the register tests the aggregate rate. 4. **Optional stopping by participants.** People stop when they are bored or losing. Since the card is committed before the guess, stopping cannot select for favourable outcomes, but per-participant rates are still reported with their trial counts so a short lucky run cannot masquerade as a high rate. 5. **Client-side timing.** No claim is made about response time; it is not collected. ## 12. What would change our minds A hit rate above 0.20 that survives the preregistered test at a milestone, with a clean card-frequency distribution, would be a positive result and published as one. The next step would be replication under a preregistered confirmatory register with an independently audited random source — not a claim about psi. ## 13. Revision history | Version | Date | Change | |---|---|---| | v1.0 | 2026-07-24 | Filed. | Once collection opens, §§2, 4, 5 and 6 are frozen.