Ground-truth data, with proof on every row.
Buy human-labeled training data — RLHF rankings, annotation, moderation, transcription — from an identity-verified workforce. Every label ships with a proof-of-work attestation your ML, audit, and procurement teams can stand behind.
From spec to verified labels in hours.
Define the task
Paste a rubric or use a 5arz template — set the label schema and acceptance criteria for your data type.
Fund it with Stripe
Set your rate and volume. You only pay for labels that pass verification — no minimum on self-serve.
Verified humans label
Identity-verified contributors complete it; every submission is server-scored for quality and fraud.
Export with provenance
Receive clean rows in your schema — CSV, JSON, or webhook — each carrying its proof-of-work trail.
Labels you can prove a human produced.
Other vendors hand you labels and ask for trust. We hand you labels with a verifiable record attached.
Verified-human guarantee
Every contributor passes Stripe Identity. The identity gate keeps the same accountable humans in the pool — not a churning anonymous crowd.
Proof on every output
Each row carries a signed attestation: anonymized contributor reference, timestamp, dwell time, anti-fraud signal, consensus score, and a proof hash.
Bot-resistant by design
Scoring is server-authoritative — the client never sets the value. Rushed-answer and dwell-time detection erode trust on low-quality work.
Every label ships with its proof.
Not just an answer — a verifiable record in your schema, ready for ML, audit, and procurement.
- ✓ The label / judgment itself, in your schema
- ✓ Verified contributor reference (identity-checked, anonymized)
- ✓ Timestamp, dwell time, and anti-fraud signal
- ✓ Proof-of-work hash + provenance trail
- ✓ Quality / consensus score where you need agreement
{
"label": "safe",
"task": "content_safety",
"contributor": "vh_4821", // identity-verified
"verified": true,
"dwell_ms": 6240,
"anti_fraud": "passed",
"consensus": 0.96,
"proof_hash": "0x9f3c…a17e",
"ts": "2026-06-25T14:02:11Z"
}
Built for teams that need defensible data.
RLHF & DPO preference data
Pairwise rankings and preference signals with provenance — so your alignment data is auditable, not anonymous.
Content moderation labels
Policy-graded safety labels from accountable humans, with a proof trail for your safety and compliance reviews.
Image annotation
Bounding boxes, segmentation masks, and hard edge cases — consensus-scored where you need agreement.
Translation QA & documents
Native-speaker translation review and structured document extraction for domain-specific products.
Pay for accepted labels.
No seat licenses, no lock-in. You only pay for labels that pass verification.
- ✓ No minimum, no sales call
- ✓ All data types
- ✓ Proof-of-work export
- ✓ Email support
- ✓ Everything in Self-serve
- ✓ Custom templates & API
- ✓ Webhook callbacks + SLA
- ✓ Provenance dashboard
- ✓ Prepay volume blocks
- ✓ MSA, DPA, data residency
- ✓ Dedicated verified cohorts
- ✓ Account executive
First-batch pilots available for AI labs and academic research groups. Ask about a free pilot.
Questions data buyers ask.
What data types can I order? +
RLHF / DPO preference data, content safety and moderation, image annotation (boxes, masks, edge cases), translation QA across 20+ languages, audio transcription and labeling, and document understanding for structured extraction. If a human can judge it, you can order it.
How do I know a real human produced each label? +
Every contributor is verified via Stripe Identity, and each row carries a proof-of-work record: anonymized verified-contributor reference, timestamp, dwell time, anti-fraud signal, consensus, and a proof hash — all exportable for independent audit.
How is pricing structured? +
You pay per accepted label. Self-serve is pay-as-you-go with no minimum; volume rates apply for recurring pipelines; enterprise prepay blocks reach the lowest per-label cost at high annual volume.
How fast are first results? +
Most teams receive a first batch within hours of submitting a spec. Paste a rubric or use a template, and connect API + webhooks to stream volumes from your training pipeline.
What stops bots or click-farms? +
Scoring is server-authoritative — the client never sets the value — with dwell-time scrutiny and trust scoring that erodes on rushed or low-quality answers. The identity gate keeps the same accountable humans in the pool.
Can I get a provenance export for compliance? +
Yes — every batch comes with a full provenance trail your procurement and audit teams can verify independently. The verification approach is covered by USPTO Provisional Patent 63/979,474.
Ground-truth data, with proof.
Create your business account and submit your first spec in under five minutes.