Every number, tested on data it was not fit on.
Five of our own headline findings were withdrawn under our own scrutiny, and every retraction is printed. The full analysis is available under NDA.
We tried to break our own findings. They held.
Every number here was tested on data it was not fit on. Five headline claims of our own were withdrawn under our own scrutiny, and every retraction is printed. A corpus whose owner kills their own headlines is worth more than one with a longer list of confident ones.
We measured what makes people pick one answer over another. Then we measured it again from scratch, on 277,759 rounds from 49,649 people none of those numbers had ever touched. The grey ring is the first measurement, the orange dot is the second. They land on the same point every time, and the biggest gap anywhere was six thousandths.
Each number is how often that kind of answer wins, measured against an average one. 1.18 means it wins 18% more often. 0.73 means it wins 27% less.
We scored each kind of signal on its own, then turned an automated search over millions of combinations loose on all of it at once. The search scored 51.6 and lost to the simple model. What is left over is not a formula we are missing. It is people being people, and the only place to get more of it is more people.
The only thing that moved it was more people. Let the room predict the verdict before it lands, feed that in, and the model gains nearly three points, measured on live rooms against their own baseline. New accuracy comes from new information, not new math on old columns. That is the whole business in one line.
On a 0 to 100 dial, the sharpest tenth of users miss someone's hidden target by 12.6 points. The weakest tenth miss by 27.0, no better than knowing nothing about the person at all. Reading people is a measurable skill with a wide range, scored against a known answer, on over 15,000 predictions made before the outcome was known.
Preference, with the loser recorded
Not just what a person chose. What it beat, in the same moment, from the same options. Most of the information in a choice lives in the options that lost, and we keep them.
A score for every judge
How well each person reads the room, scored on predictions made before the verdict. A second measure scores how well one person reads another against a known answer. The pipeline weights every judgment by the person who made it.
The reason, attached
Tens of thousands of verdicts arrive with the judge's stated why: most true, most absurd, best written, meanest, or in their own words. Preference with a rationale is a different product, and we price it as one.
Text written by the person
One answer in ten is typed by the participant, for that prompt, for that room. They beat the deck at 1.27 times chance. Every one is clean provenance by construction.
Where the line sits
Tolerance, tone and timing. How far a person will go, how far a room will let them, and the gap between what someone enjoys and what they will admit to. Both numbers held, per person, in twenty-two languages.
A personality and taste profile, built from choices, not questionnaires.
Every user's choices place them on a profile that sharpens with each answer. One dimension, how far someone is willing to go, is confirmed directly by their behavior. The others are first cuts, and a bridge study against a standard personality test is the next step. The shape is illustrative.
How far this person goes before the room turns, measured per person and per room. It is the difference between what someone enjoys and what they will admit to, and both numbers are held.
Four conditions. No dataset on the market satisfies all of them.
Nobody is compensated for an opinion. There is no incentive to produce one, so what comes out is the one they had.
No panel, no application, no vetting by other vetted people. Participants arrived on their own and stayed on their own.
Every judgment is made in front of people the participant knows, for stakes they care about. Nobody performs a laugh for a researcher.
Consent is asked at the moment of capture, scoped and revocable. Records carry it with them. Nothing here was scraped.
Rates, not totals. The corpus refills every night whether or not anyone is buying it. Two measures: one captures taste, one captures how accurately a person reads another person against an answer that is known. That second one is what makes a quality score per judge possible at all. Counts are floors, and a data sheet states the source and licensing scope of each figure.
The choice, the options it beat, the judge's score, the reason, the consent.
Two formats ship. A judgment graph keyed to option identifiers, and derived features on every option. Option text ships only where the provenance is ours, and a tenth of it is written by the participants themselves.
Illustrative shape. The feature vocabulary, its ranges, and the full schema travel with a sample under NDA.
{
"round": "r_8f3c…", "format": "fill_in", "lang": "es",
"options": [
{ "id": "c_41a…", "features": { … }, "authored": false },
{ "id": "c_9d2…", "features": { … }, "authored": true },
{ "id": "c_c07…", "features": { … }, "authored": false }
],
"verdict": "c_41a…", "reason": "most_true",
"judge": { "id": "p_ee1…", "room_read": 0.58, "n_scored": 212 },
"predictions": [ { "by": "p_2b8…", "pick": "c_c07…" }, { "by": "p_77d…", "pick": "c_41a…" } ],
"consent": { "scope": "research+commercial", "captured": true, "revocable": true }
}