RLHF & Preference Data
Pairwise judgments and reward signal from verified professionals in-domain — people whose employment records show they've done the work they're grading. Rationales documented, quality audited, identity never in question.
Rainbow staffs RLHF, evaluation, and multilingual annotation from a network of 15M+ professionals we've employed and paid across 120 countries — real roles from real contracts, validated tenure, confirmed languages. We can verify our workforce because we were the employer.
A resume is a claim.
A contract is a fact.
Most AI-training talent is vetted with a resume, a LinkedIn profile, and a contractor agreement — none of which confirms what someone actually did, how long they stayed, or whether they were any good at it. We can confirm it, because we were the employer.
Pairwise judgments and reward signal from verified professionals in-domain — people whose employment records show they've done the work they're grading. Rationales documented, quality audited, identity never in question.
Gold-standard task demonstrations authored by practitioners — diagnoses, briefs, repair procedures, financial analyses — with the reasoning models should absorb, from people we can vouch for by contract.
Model report cards graded by domain experts with verified credentials: side-by-sides, capability probes, and adversarial passes — with confidence intervals, not anecdotes.
Dozens of languages staffed by native speakers whose fluency we confirmed from where they actually lived and worked — not a self-ticked checkbox on a marketplace form.
Doctors, nurses, lawyers, insurance and financial-sector specialists, IT professionals, project managers — matched to tasks by the roles and responsibilities in their actual employment contracts.
Electricians, mechanics, CNC operators, aviation technicians, energy and agriculture specialists, defense-sector engineers — the technical and blue-collar depth most RLHF vendors can't reach through a sign-up form.
Anthropic disclosed that competing labs used roughly 24,000 fake accounts to distill Claude's outputs — an identity failure on the input side. The same failure mode sits on the human side of every RLHF pipeline: without ground-truth employment data, nobody can confirm who's actually doing the work.
The person doing the work is the person we employed — with the record to prove it.
Payroll, tax, and benefits through real entities in-country. No 1099 gray zone.
If anyone asks who's legally responsible for the workforce, the answer is us — in writing.
Dedicated operations in every region — not a single annotation role per country, but real occupational range in each.
Doctors, lawyers, and financial specialists — and the electricians, CNC operators, aviation technicians, and agronomists most vendors can't reach through a marketplace sign-up form.
“The first vendor that could tell us — with records, not assurances — exactly who graded our medical evals. Our procurement team had never seen that before.”
“We replaced a marketplace pipeline with their employed pods. Quality went up, attrition went down, and our counsel finally stopped asking about worker classification.”
We've been running this supply chain quietly behind some of the leading data-annotation companies. Now we're formalizing it — a verified, employed, compliant workforce for AI training.
Anthropic's distillation disclosure was an input-side identity failure. The same failure mode is sitting, mostly unexamined, on the human side of every RLHF pipeline.
Self-reported titles are marketing. Contracts, payroll, and tenure are evidence. The difference shows up in your training data.