Post-training
RLHF & preference data
Rubric-based ratings, pairwise preferences, reasoning traces and expert rewrites from credentialed professionals, in your target fields and languages.
For AI teams
Human data from people you can verify.
For frontier labs and AI-data vendors. RLHF and preference data, held-out evaluations, red-teaming, reasoning rubrics and multimodal ground truth, produced only by vetted experts in the domain you need, with provenance for every item.
Pilot first · Provenance for every item · 100+ languages
What we deliver
Expert data for the parts generalists get wrong.
Each project is staffed by verified practitioners in the exact field and language, and every item they produce is traceable to who made it and how it was reviewed.
Evaluations
Held-out benchmarks
New, unpublished test sets that expose where your model fails, with every error labelled and every item scored.
Safety
Red-teaming
Adversarial probes written by domain experts: clinicians, lawyers, security analysts and native speakers who know where the real risks are.
Agents & RL
RL environments & verifiable tasks
Tasks with checkable answers and reward signals, designed by the people who do the work every day.
Languages
Reasoning in 100+ languages
Native speakers with subject expertise write and grade the data, from major world languages to low-resource ones. No translated English prompts.
Ground truth
Multimodal annotation
Image, document, audio and video annotation, transcription and translation, with domain labels typed blind and adjudicated.
Who we work with
Built for the teams that buy expert data.
Frontier AI labs
Managed expert pods for post-training and evaluation
- Scoped to one capability gap at a time
- Pilot measured on your own evaluation before scaling
- Provenance record delivered with every batch
AI-data vendors
Verified specialists for fields and languages you can't staff
- Experts matched to your project's field and language
- The same five checks and records, whoever the end client is
- Finished data and evaluations, or specialists on your projects
Proof, published
We gave AI models a 1908 Urdu text. The best still got 1 word in 11 wrong.
A vetted native expert typed the truth blind, every disagreement was checked against the scan, and every word was scored. On our public leaderboard the top model, Claude Fable 5.1, reproduced 91.08% of words exactly and still made a mistake on 111 of 161 lines. Many errors carry no warning flag, so nothing in the output tells a reviewer to look twice.
- 91.08%
- best strict word accuracy (Claude Fable 5.1)
- 111/161
- lines with a mistake, even for the best model
- 2,378
- words and marks typed blind by a vetted native expert
- 8
- models ranked so far, more added as they finish
The same method runs in any field: truth written blind by a vetted expert, adjudicated against the source, scored item by item.
How verification works
Five checks before anyone touches your data.
Nobody works on a client project until all five are complete, and each result stays on record, so every item can be traced to a qualified person.
-
Identity
Government-ID verification with consent, a liveness check and duplicate-account detection.
On record: ID match
-
Credentials
Degrees, licenses and registrations checked with the issuing body: medical boards, bar associations, engineering institutions, universities.
On record: registry check
-
Domain assessment
A role-specific test, written and graded by senior practitioners in the field.
On record: score vs. pass mark
-
Paid calibration
Trial tasks scored against gold answers before any client work begins.
On record: agreement with gold
-
Continuous quality
Ongoing agreement scoring and spot audits. Anyone below the bar comes off the project.
On record: per-batch audit
See what this looks like for one expert: a sample evidence pack
How an engagement runs
From capability gap to delivered data, pilot first.
Scope
Tell us the capability gap, the domain and your quality bar. We design the task and the rubric with you.
NDA on requestMatch
We assemble a pod of verified experts whose credentials fit the task, not whoever is online.
Five checks passedPilot
A fixed-price pilot with gold tasks, dual review and agreement scores you can see.
Fixed priceDeliver & scale
Data in your format with its provenance record attached. Scale once the quality is proven.
Provenance per item
Why Vettedlance
Built for trust, not volume.
| Criterion | Generic crowd platforms | Vettedlance |
|---|---|---|
| Who does the work | Whoever is online | Practitioners matched to the task by field and language |
| Identity | Self-reported | Government-ID verification with consent |
| Credentials | CV claims | Checked with the issuing body |
| Quality control | Spot checks after delivery | Gold tasks, calibration and agreement scores on every batch |
| Provenance | Rarely documented | A record for every item: who, when, which guideline, how reviewed |
| Coverage | Generalists, English-first | Recruited by domain, in 100+ languages |
| Contributor terms | Opaque, variable | Written contract, rate agreed before work starts |
Questions
Questions, answered plainly.
Can we start with a pilot?
Yes. Most teams start with a fixed-price pilot, then scale once quality is proven.
How do you prove data quality?
Every item carries provenance: which vetted expert produced it, when, under which guideline version, and how it was reviewed. We share our method and sample results openly.
Which languages and fields?
We recruit and vet experts across medicine and pharma, law, the sciences, engineering including aerospace, software, finance, defence, education, the arts, and more than 100 languages. We staff each project by demand. Every expert on a project has passed identity, credential and skill checks.
Can you sign an NDA before we share details?
Yes. Tick the NDA box when you request a proposal and we will send one before any project details are exchanged.
Are you already working with AI labs?
We're onboarding our founding expert cohort and building our first datasets and benchmarks. We don't list clients or partners we haven't signed, and we never will.
Get started
Tell us the capability gap. We'll scope the pilot.
Share the domain, languages and quality bar. We reply within two business days.
For AI labs & data vendors
Scope a pilot in your domain, languages and quality bar. We reply within two business days.
For experts
Paid, remote AI work in your field. Free to apply, with a written contract for every project.