For Builders

Annotation, as protocol.

Verified labels. Signed datasets. Direct payment to Humans.

§ 02 — Why us

Why Builders/
pick Humaniti

  1. 01
    Verified at the protocol layer.
    Every label routes through N-Human consensus. Disagreement triggers adjudication. You don't audit quality — the network does.
  2. 02
    Pay per verified label, not per hour.
    Pricing is transparent and per-task. No setup fees, no minimums, no net-60 invoices.
  3. 03
    Datasets you own.
    Outputs are signed and addressable. Export anytime in 16+ formats. The dataset lives outside our database.
  4. 04
    API + SDKs.Early access
    REST endpoints plus a Python and TypeScript SDK in early access — full reference docs at general availability. Request early access →
§ 03 — Coverage

Cover every modality.

17 task types across 5 modalities. Tooling tuned per modality. Throughput scaled by the qualified pool.

Pixel-precise labeling for image and frame data.

Image bbox

Tight rectangles around objects. Multi-class, multi-instance.

COCO·YOLO·Pascal VOC

Polygon

Multi-vertex outlines for irregular shapes. Higher precision than bboxes when objects do not align to axes.

COCO·GeoJSON

Segmentation

Pixel-level masks per object or class. Semantic, instance, or panoptic.

PNG masks·COCO RLE

Keypoint

Named points on a body, face, or pose skeleton. 17-point COCO body, 68-point face, custom skeletons.

COCO
§ 04 — Quality

How quality stays high.

  1. 01
    N-Human consensus.
    Every label is independently produced by N Humans. Default N is 3. High-stakes work runs at 5 or 7.
  2. 02
    Elo + Tier system.
    Per-skill Elo + global composite. Verified accuracy raises tier. Tier gates which work a Human can pick up.
  3. 03
    Gold questions.
    Known-answer items injected into every batch. Continuous calibration without telling the Human which is which.
  4. 04
    Dispute flow.
    Disagreement triggers blind re-annotation. Steward adjudicates. Outcome posts to the audit log.
  5. 05
    Stewards.
    T3 Humans who adjudicate disputes, audit gold-question performance, and review edge cases.
  6. 06
    Anti-gaming.
    Behavioral telemetry, text-detection ensemble, device + wallet clustering. Slashing lands in Phase 3.
Read the full methodology →
§ 05 — Integration

Three ways to start.

Web wizard.

Drag-drop a sample. Pick a task type. Set N-Human consensus. Publish. No code, no install. The shortest path from sample to verified labels.

§ 06 — Pricing

What it costs.

Per-label rate, multiplied by your chosen N-Human consensus, multiplied by volume, plus a transparent network fee. No setup, no minimums, no quotes. The full grid is on the pricing page.

See full pricing →

Apply for a workspace.

Approval typically inside 2 business days. We review every application.