PHD (Physical Data) is a physical-world data factory producing spec-driven robotics datasets at scale: human demonstrations (imitation learning), manipulation sequences, QA-gated releases, and custom schemas for VLA pretraining & policy learning.