How a label becomes verified.

From request to receipt.

A Builder defines the task — its prompt, its label schema, its consensus N, its quality thresholds — and publishes. The qualified pool is filtered by tier, skill, language, and region. Each item routes to N Humans independently. The aggregator weighs each label by Elo, accounts for known per-Human bias, and returns a consensus label plus a confidence score.

If confidence is high, the label ships into the signed dataset and the Humans get paid. If confidence is low or a Human appeals, the item routes to a Steward. The Steward's outcome is final and writes to the append-only audit log.

Quality is a ladder.

On top of the four tiers, each Human carries a per-skill Elo — bbox, Text NER, audio transcription, RLHF preference, and so on. New Humans seed their Elo through a qualification arena of gold-graded items. Verified accuracy raises the score. Drift lowers it. Higher Elo unlocks higher-paying work in that specific skill.

What a Builder controls.

1. 01
   N-Human consensus.

   1 (cheap-draft) / 3 (default) / 5 (high-stakes) / 7 (max). Cost scales linearly with N. Confidence scales nonlinearly.
2. 02
   Gold-question density.

   Percent of items in each batch that are known-answer plants. Higher density tightens calibration; lower density saves cost.
3. 03
   Target agreement threshold.

   The inter-Human agreement (IAA) score below which a label routes to dispute. Tunable per project. Defaults tuned per task type.
4. 04
   Language and region constraints.

   Restrict the qualified pool by language, region, or device. Required for locale-sensitive work.
5. 05
   Steward escalation thresholds.

   Disagreement count, time-on-task, gold-question failure rate — any of these can trip Steward review. All defaults configurable.

Disagreement is data.

Plain majority vote is a bad aggregator. It treats every Human as equally accurate and every disagreement as noise. The Network treats disagreement as signal. Three families of methods do the work — Dawid-Skene estimates each Human's confusion matrix from their full label history and weighs their vote accordingly. MACE separates random spamming from honest disagreement. CAZ peer-prediction rewards Humans for being informative about what other Humans will say, which makes random clicking a losing strategy.

The aggregator returns two numbers — the consensus label and a confidence score. The confidence score is what trips dispute, gold-question audits, and Steward escalation. The math stays inside the Network. The Builder sees the label, the confidence, the audit trail, and the receipts.

Why this can't be cheated.

• Behavioral telemetry — keystroke timing, pointer paths, dwell, scroll. Anomalies route to review.
• Text-detection ensemble — multi-model classification of generated text in free-response fields.
• Gold-question injection — known-answer items mixed continuously into every batch, blind to the Human.
• Wallet and device clustering — fingerprint and payout-rail signals collapse Sybil rings.
• Steward audits — T3 Humans periodically re-grade samples and review flagged sessions.
• Phase 3 slashing — staked credits are slashable for confirmed gaming once the credit unit ships.

What happens when Humans disagree.

1. 01
   Mandatory rejection reason.

   Any Human who disagrees with a peer must record the reason. Free-text plus a structured taxonomy.
2. 02
   72-hour appeal window.

   The original Human is notified. They can accept, revise, or escalate within 72 hours. Silence accepts the consensus.
3. 03
   Blind re-annotation.

   If escalated, the item routes to a fresh pool of Humans who see the prompt without prior labels or rejection notes.
4. 04
   Steward adjudication.

   A T3 Steward reviews the item, the consensus history, and the appeal. The outcome is final and posts to the audit log.