This working paper describes a systems pattern implemented in EvoMind for converting a high-level user objective into one or more real, verified work products without granting the planner direct execution authority. The pattern separates semantic planning, capability resolution, governance admission, execution, independent verification, and artifact-state admission. In the current implementation, a high-level objective can be classified as requiring no deliverable or one or more deliverables; a CapabilityGoalGraph expresses work-product requirements and dependencies; a registry-grounded capability broker selects implementations; canonical admission and governed execution control side effects; and verified artifact records preserve identity, versioning, provenance, and lineage. An internal live acceptance proof through the conversational handle_chat() path produced real Word, Excel, and PowerPoint artifacts from a single high-level goal while an ordinary informational question produced zero files. The frozen acceptance matrix passed 9/9, with a 252/252 surrounding regression set. These are internal engineering results, not third-party replication. The paper positions semantic capability composition as an architectural layer between intent and tool execution, and argues that trustworthy computer-use systems benefit from separating 'what must become true' from 'which capability should act' and from admitting outcomes into trusted state only after verification. The paper does not claim universal desktop autonomy, formal correctness, or artificial general intelligence.