A worker kneeling beside a solar panel is doing ordinary construction work. But to a robot, successfully identifying that worker may present an edge case: the worker is low to the ground, partly hidden by equipment, and wearing heavy PPE. The problem in identifying the worker starts before model training. Annotation pipelines often reject images when they can’t confidently reconstruct a worker’s full pose. That may improve the dataset's apparent quality, while removing the difficult examples a safety-focused system most needs to learn from. At Gritt, we call this the selection-scarcity confound. After auditing four months of construction-site data, we found that low poses accounted for just 2.2% of the retained person instances. Further analysis suggested that this scarcity was partly created by the filtering process itself. So we separated person detection from pose-label quality, then brought the missing postures into simulation. Adding simulated workers who were kneeling, squatting, or deeply bent over improved detection recall on real site footage for those poses by 1.6 percent and low pose detection by 5 percent. In physical AI, rare examples aren’t necessarily unimportant... in fact, they may be the *most* important when it comes to safety. Read the full post from Gritt Perception Lead, @r15hav, with a link to the paper on arxiv: gritt.ai/news/the-safety-cas…