How Trates works.

Methodology version: trates-1.0. Trates separates what participants value from how participating friends experience one another.

Behavioral observations

Each of 96 curated traits has a stable definition, a behavioral prompt, and a version. Answers use a five-point frequency scale from rarely to consistently. “I haven’t seen enough to say” contributes no score. Each accepted friend can hold one current answer per trait per monthly round. Self-reflections are stored separately.

Private portraits

At least five distinct accepted friends must contribute eligible non-skipped observations for a dimension. The released band is the rounded mean on the five-point scale, with an indication when experiences vary. Dimensions are frozen once released for that monthly round. Missing dimensions are withheld rather than treated as low scores.

Eligible first dimensions release on Sundays at 12:00 UTC through a scheduled job. Results do not update answer by answer. A failed job retains the last valid snapshot. Privacy changes may withdraw affected results until a later eligible release.

Trait preferences

Official matchups offer two traits, a tie, and a skip, separately for general desirability, friendship, partnership, family, and work. The server issues balanced-exposure pairs and bounds contributions to 30 per day and one vote per person, pair, context, and month.

Rankings fit a Bradley–Terry logistic preference model with L2 regularization of 1. Ties contribute one half to each side; skips are excluded. Publication requires at least 100 distinct contributing people, at least 20 distinct people for every included trait, and a connected comparison graph. Weights from the preceding completed month are frozen for the current month, including when data is insufficient.

Pairwise comparison is inspired by Chatbot Arena. That work does not validate personality measurement by Trates.

People discovery

Participation is optional. Each candidate requires at least ten distinct peer contributors in the selected ranking and at least five contributors per included dimension. Overall and relationship boards use the 12 core dimensions and the corresponding frozen preference weights. A single-trait board uses only that trait and replaces relationship weighting.

Frequency averages are normalized to 0–1 and conservatively shrunk toward 0.5 with five prior-equivalent observations. Ranking combines those averages, never total votes or friend count. A board needs 100 eligible people after every filter and shows only its top ten aliases, avatars, and positive highlights. Raw scores and low dimensions are not public.

Regions

Country and state or province are optional and self-declared. Regional feedback summaries weight contributing people equally and require at least 100 eligible people and 100 distinct raters. Regional preference rankings use only consenting participants in that region and the same comparison-coverage gates. Thresholds apply after every filter.

These are Trates samples, not representative surveys of a region. Recruitment, who knows whom, social context, response styles, collusion, and missing observations can all affect results. We do not claim a state or country has an objectively more honest or optimistic population.

Changes and corrections

Monthly recaps report supported patterns and differences in comparable feedback. They do not claim that changing feedback proves personal growth. Private usefulness reactions never alter scores. Blocked, deleted, suspended, or consent-ineligible contributions trigger invalidation. Suspected manipulation can pause publication pending review.