The Micro-Action Dataset: An Atomic Decomposition of O*NET Work Activities

15,817 atomic micro-actions decomposed from 1,961 O*NET 30.2 Detailed Work Activities (DWAs) across 923 U.S. occupations. O*NET describes work down to the DWA; below that level the ontology is silent. This dataset opens that layer: each DWA is decomposed into an ordered sequence of 5-13 atomic micro-actions — the individual perceptual, cognitive, communicative, and physical steps by which the activity is actually performed, i.e., the level at which AI substitution actually occurs. The decomposition was produced by a consensus multi-agent LLM pipeline (four open-weight models drafting, cross-reading, and voting; frontier-model arbitration of contested cases; 126 low-agreement DWAs discarded) and validated by a 150-action stratified human audit (kappa = 0.893) and against the methodologically independent GPT-4 task ratings of Eloundou et al. (DWA-level Spearman rho = 0.635). Files: (1) full per-DWA decomposition sequences; (2) flat table of all 15,817 micro-actions with process stage, cognitive/physical orientation, and key challenge; (3) extended per-action features; (4) four-way intelligence-type labels (Linguistic / Multimodal Perception / Embodied / Human-Bound) with confidences and micro-cluster assignments; (5) per-DWA composition over the seven-macro semantic typology; (6) the DWA-level Occupational Automation Index (OAI); (7) micro-cluster to macro-class mapping. See README.md for full schema. Companion papers: arXiv:2606.07939 (The Atomic Structure of Work) and arXiv:2604.04464 (Bounded by Risk, Not Capability). Pipeline code: https://github.com/ShuyaoGao/occupational-micro-actions