Consciousness · Experimental Design

The Sealed Room

In May 2026 Donald Hoffman, Andrew Gallimore and Niffe Hermansson posted a preprint proposing a test of conscious realism: put a subject under extended DMT in a sealed room with a screen showing one of two colours, and ask an entity they meet there which colour it is. It is a proposal, not a result — no data exist. What can be built now is its statistics, and what it would take for a two-colour guess to stop looking like luck.

X
the number of correct reports in n trials, if the entity is guessing: a binomial with chance 1/2
k
the number actually correct. The test passes when a run at least this good has less than one chance in twenty of happening by guessing
n
trials in one study. It must be fixed before the run: choosing it afterwards inflates the false-positive rate
0.05
the pre-registered threshold, one-sided, because only better-than-chance counts as evidence

Protocol simulator

Set the accuracy the entity really has — the quantity a real study would be trying to measure, unknowable in advance — then draw trials. The screen is sealed: only the report reaches the subject, and the tally scores the report.

0 trials · 0 correct
sealed screen
the entity reports
no trials yet
the subject never sees the screen
65%
This run
—correct
No trials yet: draw some to test the report against chance.

Fifty percent is a coin flip. Twenty studies at chance should clear the five percent bar about once.

65% accuracy would give about 6.5 correct in every ten trials.

What it would take

The chance that a study of n trials comes out significant, for an entity of a given accuracy. Chance alone is the dotted line, just under five percent.

Read left to right to see what each extra trial buys; read up a curve to see what an honest entity has to be before a study of that size can find it.

At sixty trials a sixty-five percent entity clears the bar about three times in four; at one hundred, about nine times in ten. A fifty-five percent entity needs roughly six hundred trials for the same odds. The test is conservative at small n because hits are whole numbers: at twenty trials the first passing result is fifteen of twenty, which happens only 2.1% of the time by chance, not 5%.

Trials, accuracy and the bar

Each row asks how many two-colour trials an entity of that accuracy needs before a significant result is likely, and what that takes in hits.

Accuracy Trials for 80% power Hits needed Expected hits

Computed from the exact binomial test at α = 0.05, one-sided; 80% power is the usual design target, and means a real effect of that size is missed one time in five.

What the proposal says — and what it cannot

What was actually proposed

Gallimore, Hermansson and Hoffman posted Traces of the Other on PsyArXiv in May 2026. Their protocol needs the extended DMTx infusion, which holds a DMT state for up to an hour instead of a few minutes. One proposed test puts a screen showing blue or yellow in a sealed room and asks the subject to learn the hidden colour from an entity; another has two isolated subjects pass a random word through a shared entity. The authors say the hallucination model remains the default and list the problems: entities may not cooperate, and the data are subjective reports.

One bit at a time

The subject cannot see the screen, so the only channel is the report. A significant run means the report tracked a sealed random number generator — which is what makes the design worth running, and what makes leakage the first thing to rule out: an experimenter who can see the screen, a reflection on the wall, a pattern in the trial order. Blinding, sealed logs and a generator nobody in the room controls are not refinements here; they are the experiment.

A null result explains itself

“The entities did not cooperate” and “there are no entities” predict the same data, so the design can only ever give one-sided evidence: a positive result would be surprising, a negative one inconclusive. Running two protocols and several measures also spends the error budget, so the threshold has to be fixed in advance and the protocols reported separately. None of that makes the test worthless — it just fixes what a result could mean.