# EXP-002 — Generic versus specific personality descriptions **Status:** filed, collection not yet open **Protocol version:** v1.0 **Filed:** 2026-07-24 **Experiment slug:** `forer-machine` This document is the preregistration. It is published before any data exists, and its SHA-256 is displayed on the experiment page so a reader can verify that the plan shown to them is the plan that was filed. --- ## 1. Question People rate vague, universally applicable personality descriptions as strikingly accurate accounts of themselves — the Barnum or Forer effect, after Forer (1949). Demonstrations of it are common; controlled measurements of it are less so. The usual demonstration has no comparison group: everyone reads one generic text, rates it highly, and the reveal is the whole content. That shows something happened but cannot say what, because there is nothing to weigh the rating against. A mean of 7.8/10 is only interesting next to what people give a description that is *not* built to fit everyone. So this register is a two-arm comparison, not a demonstration. ## 2. Design Each participant is assigned to one of two arms and rates a single description for accuracy on a 1–10 scale. - **Arm A — generic.** Ten statements written to apply to nearly anyone: double-sided, mildly flattering, unfalsifiable. - **Arm B — specific.** Ten statements of the same length, register, and sentence rhythm, but making concrete claims that cannot be true of most people (sleep patterns, attitude to animals, comfort with numbers). Both texts are fixed, identical within their arm, and published in full in the site repository. Neither is generated per participant, and no participant's input influences the text they receive. **Assignment** is deterministic on a hash of the participant identifier, so a reload cannot reassign anyone, and the split is even without server state. ## 3. Hypotheses - **H₀ (null):** mean accuracy rating is equal in both arms — the difference is zero. - **H₁:** Arm A is rated higher than Arm B. ## 4. Primary outcome Mean self-rated accuracy (1–10) per arm, and the difference between arms. ## 5. Statistical analysis Fixed before collection, not modifiable while the register is open: - **Test:** Welch's two-sample t-test on mean rating, two-sided, α = .05. Welch rather than Student's because equal variances between arms are not assumed. - **Effect size:** Cohen's d with a 95% confidence interval. - **Robustness:** Mann–Whitney U reported alongside, since a 1–10 rating is ordinal and may be skewed. If the two disagree, both are reported and neither is presented as the "real" result. - **Milestone reads:** at n = 1,000, 5,000, and 10,000 total entries. - **Interval estimates:** 95%, normal approximation for means. ## 6. The birth-data step, and why it is theatre Before the description appears, the participant is asked for a birth date, time, and place — the ritual that usually precedes a "personalised" reading. **That data is never transmitted.** It is held in the browser, used only to render a moment of anticipation, and discarded when the page closes. No birth data reaches the server, appears in the database, or exists in the export. The reveal states this plainly and invites the participant to check their browser's network tab, where they will find no request carrying it. This is the minimum deception the effect requires. Participants are told truthfully at the outset that they are taking part in an experiment about how people judge descriptions of themselves; what is withheld is only that the text is fixed and that a second arm exists. Both are disclosed immediately after the rating is submitted, in the same page view — there is no debrief the participant has to come back for. No participant is told that a description was computed from their birth data. The page never claims it. ## 7. Procedure 1. Participant reads what the experiment is about and enters birth details. 2. A description appears, drawn from their assigned arm. 3. They rate its accuracy, 1–10. 4. The rating is recorded, and the reveal follows immediately: both texts, which arm they were in, the live distribution of ratings for each arm, and where their own rating sits. ## 8. Eligibility and deduplication One rating per participant, enforced by a database constraint. A participant who has already rated sees the reveal rather than the form. Clearing browser storage mints a new identifier; as in EXP-001, the count is not claimed to be tamper-proof, and any exclusion rule applied in analysis will be published with the data. ## 9. Anchoring The rating distribution is not readable by the page until a rating has been submitted: it is returned by the same call that records the entry. Nobody sees the crowd before judging. ## 10. Data publication Anonymised entries — arm, rating, referral source, date — are published and refreshed from the day collection opens, under CC BY 4.0. No birth data, because none is collected. ## 11. Prespecified threats to validity 1. **Self-selection.** Participants arrive from communities with priors about astrology and about psychology. Referral source is recorded so the split can be inspected. Results describe this sample. 2. **Arm B may be rated low for the wrong reason.** If its statements read as oddly specific rather than merely inaccurate, participants may rate it down for style rather than fit. Mitigated by matching length, register and rhythm across arms; the texts are published so the matching can be judged. 3. **Demand characteristics.** A visitor to a site called Occult Research may guess that a trick is coming and rate defensively. This biases toward the null, which makes a positive difference the more robust outcome. 4. **Single pair of texts.** Any difference is a property of these two texts as much as of the categories they represent. A replication with new texts is the correct way to test whether it generalises. 5. **Ordinal ratings treated as interval.** Addressed by reporting Mann–Whitney alongside the t-test (§5). ## 12. What would change our minds If Arm A and Arm B are rated equivalently at a milestone, the Barnum effect did not appear under these conditions, and that is published as plainly as a positive result would be. It would most likely mean the texts were poorly matched, and the next step would be new texts — stated here, before the data exists, so it cannot be read as an excuse invented afterwards. ## 13. Revision history | Version | Date | Change | |---|---|---| | v1.0 | 2026-07-24 | Filed. | Once collection opens, §§2–5 are frozen: no revision may alter the design, the hypotheses, the test, α, or the milestone schedule while the register is open.