Abstract: Geometric constraints are used by many sketch recognition systems to perform high-level assembly of components of a sketch into semantic structures. However, with a few notable exceptions, most of the current recognition systems do not have constraints that use real-valued notions of confidence. We discuss methods for assigning confidence values to different kinds of constraints. We show how these confidence values equate to user perception, how they can be used to balance speed and accuracy in recognition algorithms, and how they can be used to assign confidence values to the high-level shapes they are used to construct. We use these constraints to extend the LADDER shape definition language in a system that recognizes 5,900 hand-drawn examples of 485 different military course-of-action diagrams at an accuracy of 89.9%.