Join the Testing Pool
Keep your reading with an account, return to it on any device, and take part in the research once you finish signup.
Calibration
Give one private 1 to 7 attraction rating at a time. Your answer is saved only after the page confirms it.
Your progress
Progress follows your account and remains separate from later research results.
Appearance information
This private information is shown only when the current evidence and every release check allow it.
What it is
Most dating apps match on profiles. Harmonia is testing whether what you are drawn to, how you respond to real situations and, for some people, your biology say more. No other platform has combined these three signals. Three planned phases test it with real people.
Visual preferences and personality mapping
Rate faces from 1 to 7 and answer six everyday situation questions in your own words.
Your ratings start from the crowd's view of each face, and the system learns how yours differ. A reading appears only once enough ratings stand behind it, in wording written in advance.
Mutual matching and perceived similarity
Choose between two people at a time, each open to the other as you are to them.
Each side's likely interest is estimated separately, so a pair is considered only when both directions point the same way.
Biological compatibility via HLA analysis
By invitation, take a free DNA test, or share results you already have.
The research looks at whether HLA fit, beside what you are drawn to and how you respond to situations, says more about who you would want to meet. This phase opens later, by invitation.
The phases are planned and open in stages. Anyone who completes signup can take part; only Chemistry is by invitation. Taking part does not guarantee an introduction or a match.
What has been built
Harmonia's matching system is built. It has been tested on ratings from public datasets, from See Your Eye and in simulation, not yet with Testing Pool members. Joining is how that starts.
It begins with how people in general rate each face, then learns how your ratings differ.
People are considered only when each of you is open to the other.
A reading appears only once enough ratings stand behind it, in wording fixed in advance.
Six everyday situation questions, none asking you to describe yourself.
A model that learned how each person scores predicted that person's later ratings better than the crowd average, on all three rating datasets tried.
Three rating datasets, two public and one from See Your Eye. In the largest public set, 2,462 photographs and 68,868 ratings, its prediction error fell below the crowd average's by 0.195 after 5 ratings and 0.279 after 50.
Ratings the person has given
After 5 ratings: 69% of the improvement it reaches at 100.
In simulated pools, about 1 person in 200 was told something that overstated the truth, against an allowance of 1 in 20.
200 simulated pools of 240 raters each: 9 overstated statements in 1,800 cases. A real planted effect was reported in about 1 pool in 10 at 25 ratings and 9 in 10 at 200, so it speaks up as the evidence grows.
We reran a published method for predicting each person's ratings from its authors' own code, and our runs beat its reported scores. Our simpler correction then beat its per-person tuning on average error in every test.
The paper's own benchmark, rerun from its authors' code: 5,500 faces and 20 test raters. Both versions of the method scored above the paper's reported correlations in all five repeat test runs, by up to 0.01. Given 50 of a rater's scores, its per-person tuning missed the rest by 0.562 and 0.580 points on average, on a 1 to 5 scale; our simpler correction missed by 0.533 and 0.534.
Correlation with each rater's scores, higher is better
Average error for version T, lower is better
When Harmonia's raters scored the same face again, 9 in 10 landed within a point of their first answer, and 6 in 10 gave exactly the same score.
403 repeat pairs on See Your Eye, 20 September to 1 October: 63 percent the same score, 90 percent within one point. Steady scores are what make it possible to learn how someone rates.
In simulated pools, ranking by both people's likely interest beat ranking by either person alone, in every pool type tested.
14 simulated pool types, 30 pools each. Each side is scored without the pair's own answers. The bars show ranking quality in the baseline pool (NDCG at 10).
These are research results about ratings and simulated pools, not about Testing Pool members, and not a prediction of whom you will find attractive.
Questions
Matching is part of the research, in Resonance. Taking part does not guarantee an introduction or a match.
Joining costs nothing. There is no payment or prize, and none is promised.
One main selfie, saved privately. Staff check its technical suitability. We do not assess attractiveness. Saved photographs expire after 500 days, or earlier when you remove them or request account erasure.
Yes, with your eye-print code on See Your Eye. Deleting permanently removes the reading and every rating behind it. It cannot be undone, and your code stops working.
Adults aged 18 or over. If you live in Greater London you can also register interest in possible later research; it records interest only.
Staff review your selfie's technical suitability. No review time is promised. Return with your email code and refresh your status. You then take part as the research phases open; only the DNA phase is by invitation.
Know someone who would like to see their own Eye?
Preferred name and email only. Notify Me does not create Testing Pool membership or research consent.
We use AI as a translator, never as a determinant. You are the decider.