Human preference. Measured at scale.
A model trained on our data beats every frontier model at predicting what people choose, even on content it never saw.
Each dot is ten thousand preference pairs
Frontier models misread people. Our data doesn't.
“Which option will this person pick?” On content never seen in training
frontier models beaten at predicting what real people choose, even on content our data never saw
points better at predicting people when Claude Opus or GPT-6 is given what our data taught a model
right at picking the answer people prefer, as a judge trained on our data. GPT-6 65.0%, Claude Opus 62.8%
Pre-registered tests · 95% confidence ranges · full results in the data room
Ask any population anything.
We design a gamified survey, run it with real users, and deliver labeled data. First data in two weeks.
A profile built from choices, not questionnaires.
What people choose, and what it predicts they will like next.
How well people predict what others will choose, scored against a right answer.
How people read tone, and how they react to it.
Where the line sits for different people, and how it moves with the room.
Where populations stand, inferred from choices and reported only in aggregate.
Also measurable: humor, brand affinity, trust, honesty under pressure
Before a model can move a population, someone has to measure what moves people. That is where OREX comes in.
Built the audience. Shipped AI at Amazon.

40M+ recorded choices, with zero paid acquisition.
Designed the identity, consent, and measurement architecture that turns choices into structured preference data. Patent pending on behavior-driven consent and identity capture. Previously led growth and product at Google, Lyft, and Robinhood.

Shipped production AI at Amazon before it was a category.
Built multi-model orchestration inside Alexa in 2017, routing between competing models to choose the next step in a conversation. 25+ years across Amazon and Microsoft. 9 patents.
Every number tested on data it was not fit on. Five of our own findings retracted.
See the evidenceThe human signal models are missing.
We publish our methods. Bring your hardest ML lead.
Choices with every alternative kept, a score for each judge, and the stated reason.
Any population, any question. First data in two weeks.
Score how a model reads people against a live human baseline.