https://synjuku.ai/blog/ Measured findings on embodied-AI data quality, capture, QC, and deployment, from the team running the loop in production. en Sat, 01 Aug 2026 12:00:00 GMT https://synjuku.ai/blog/caption-audit https://synjuku.ai/blog/caption-audit Sat, 01 Aug 2026 12:00:00 GMT Everyone is benchmarking whether AI can write video labels. We asked whether the labels already shipping in embodied-AI datasets are actually true. On a 136-hour egocentric manipulation dataset, roughly a third of captions failed a human check, the errors concentrate in the verbs, and a carefully calibrated VLM judge catches about half of them with 90% precision, for $3 per hour of footage.