Experiment I · protocol

The Blind Horoscope Test

Twelve horoscopes rise each morning with their signs effaced. If the stars write anything personal into them, their readers should find their own at better than one part in twelve.

EXP-001 · STATUS: Protocol published · collection not yet open · v1.2

This is a protocol, not a result

Collection has not opened, so there is no data on this page and no counter to watch. What is published here is the plan — filed in advance precisely so it cannot be adjusted once results start arriving. The charts below describe the null hypothesis, computed from the design itself.

The wager

Readers shown all twelve of a day's horoscopes — unlabelled and shuffled — will select the one written for their own sun sign more often than chance. Chance is one in twelve: 8.33%. The astrological claim predicts more; the null predicts exactly that, forever.

The analysis is fixed in advance: a two-sided exact binomial test against p₀ = 1/12 at α = .05, with conclusions drawn only at sample sizes of 1,000, 5,000, 10,000 and 25,000. Because this is filed before any data exists, there is no room to move the goalposts once results arrive.

Fig. I — What chance looks likep₀ = 1/12 · 95% · POWER 80%

Both curves are computed from the design, not observed. As the sample grows, the band of luck narrows and the smallest resolvable effect falls toward the chance line — which is exactly why the milestone schedule exists. A genuine effect holds its level while the band sinks away beneath it.

Sample size

Chart of the null hypothesis: the 95% band of hit rates consistent with chance narrows as sample size grows, and the smallest detectable true rate falls from about 15% at n=1,000 toward 9% at n=25,000.
Sample size	Band of chance	Smallest detectable
1,000	6.6% – 10.0%	10.8%
5,000	7.6% – 9.1%	9.4%
10,000	7.8% – 8.9%	9.1%
25,000	8.0% – 8.7%	8.8%

The instrument

Source: One nationally syndicated daily column, named in the protocol. The same source every day of the run — switching sources mid-experiment would quietly change what is being tested.
Blinding: Signs are stripped server-side and presentation order is shuffled per reader, so position carries no information.
Entry: One observation per participant per day, deduplicated by account and hashed IP.
Sealing: Same-day selections stay hidden until 00:00 UTC. Nobody should be able to see the crowd before choosing.
Registration: Protocol v1.2, filed before collection opens. Hypothesis, test, α, and milestones are locked; changes are appended and logged, never overwritten.
Open records: Anonymised raw data and the analysis code will be published and refreshed nightly from the day collection opens.

What a reader will see

Twelve readings from the same day, signs removed, order shuffled. The participant picks the one that matched the day they actually had — and only then is asked for their sun sign. The examples below show the format; they are illustrative text, not a live draw.

Example format — not a live draw

A conversation you've been putting off turns out easier than expected. Momentum from the morning carries one stalled task over the line.

Money asks for a second look — a small recurring cost has outlived its usefulness. An old friend resurfaces with an unusual request.

You've been carrying more than your share, and today someone finally notices. A piece of overdue good news arrives by evening.

The list is shorter than it looks. Start with the item you've circled for days and the rest fall in line.

Known ways this could mislead

01
Twelve signs means twelve comparisons. At the usual threshold, roughly one sign will look remarkable by luck alone — so per-sign results will carry a multiplicity correction, and the uncorrected figures will be shown beside them rather than quietly dropped.
02
A single well-written reading tends to attract picks from everyone regardless of sign. That is the Barnum effect showing up as a crowd favourite, and it is measured separately rather than mistaken for a signal.
03
Self-reported sun signs are unverified. Misreporting adds noise, which biases toward finding nothing — worth stating plainly, because it means a null result is the weaker of the two possible conclusions.
04
Readers who arrive from astrology communities and readers who arrive from skeptic communities differ in ways we cannot control. Referral source is recorded so the split can be inspected rather than assumed away.

What a sample can and cannot see

Smallest detectable hit rate by sample sizep₀ = 8.33% · 80% POWER

Sample size	Smallest detectable rate	Lift over chance
1,000	10.8%	+2.4%
5,000	9.4%	+1.1%
10,000	9.1%	+0.8%
25,000	8.8%	+0.5%

Sample size sets the smallest real effect an experiment could notice. These are computed for the blind horoscope design against its 8.33% chance baseline, at 80% power — not illustrations. — n ≈ 2,271 for a true rate of 10%.

Watch this register

The register opens when the data layer is live. Until then the protocol stands as published.