for AI labs
Fresh human chat is the rarest data.
Models can ace the bar exam and still sound robotic and over-eager in open conversation. Closing that gap takes real people being real: on purpose, with consent. RUB mints that signal every day.
What the park produces.
“Which reply is better?” verdicts on warmth, naturalness, and over-eagerness vs. genuine support, each with a written rationale. The lead signal.
Live adversarial naturalness labels from real chats, plus what gave the bot away, in the player's own words.
Consented human dialogue minted from play, not paid role-play. Fresh every day, scrubbed before export.
Rights-clean by construction.
Built so you can train on it and document it. Every row is something a player agreed to, and can walk back; the Privacy Policy is the canonical version of everything below.
Consented, 18+
Contributors opt in knowingly. Consent is versioned per contributor, and opting out later is self-service.
Opt-outs honored
Consent is rechecked at delivery time, never cached. A withdrawal blocks that contributor from every export not yet delivered.
Two-stage PII scrub
Conversation text passes pattern redaction plus an LLM verifier before it can ship. Short label fields get the pattern pass.
Friend DMs never ship
Private friend conversations are excluded from every dataset, categorically.
Cohorts, not identities
Demographic context ships as coarse buckets only: age band, gender, country. Rare combinations are suppressed. No identifiers.
Profiles are doubly opt-in
Participant profile context, rounded personality scores and ranked values, ships only under its own separate consent. Raw answers never leave.
Audited exports
Every full export is versioned, checksummed, and traceable back through the pipeline that produced it.
Design partners
Want the sample before anyone else?
Private beta: we're after design partners, not customers yet. React to a real de-identified sample under NDA and tell us whether it clears your bar. The no is as useful as the yes.