Quality Assurance

Data quality is the foundation of Humanoid Network's value proposition. Every submission passes through multiple validation layers before earning rewards.

Automated QA

The primary QA layer is fully automated, running ML models and heuristic checks against each submission.

Quality Dimensions

Dimension	What It Measures	Method
Completeness	Full task execution	Action segmentation model
Visibility	Hands/objects in frame	Object detection + occlusion check
Stability	Camera steadiness	Optical flow variance
Lighting	Scene illumination	Histogram analysis
Duration	Meets minimum length	Timestamp comparison
Pose Quality	Keypoint confidence	Pose estimation confidence scores

Scoring

Each dimension produces a normalized score (0-1). The aggregate quality score determines: - Pass/Fail: Minimum threshold for acceptance - Reward multiplier: Higher quality = proportionally higher rewards - Contributor reputation: Rolling average of recent quality scores

Anti-Sybil Protection

To prevent gaming and low-effort submissions: - Device fingerprinting: One account per device - Progressive verification: New contributors start with simpler quests - Similarity detection: Duplicate or near-duplicate submissions are flagged - Stake-weighted reputation: veHAN stakers have more to lose from bad submissions

Diversity Scoring

Beyond individual quality, the network scores submissions for diversity across: - Environment types: Indoor, outdoor, kitchen, office, warehouse - Body types: Range of heights, builds, and physical abilities - Geographic regions: Global coverage prevents training bias - Task variations: Multiple approaches to the same task