Lazy evaluation is a best-first state space search method for contextual handwriting recognition which searches an ordered space and applies constraints at the earliest possible opportunity to maximize computational efficiency. Lazy evaluation is well-suited to multi-level hypothesis verification because it operates recursively on a hierarchical semantic tree of context constraints. This paper describes the lazy evaluation algorithm in detail, and proves that, if ordered hypotheses for the lowest-level sub-patterns are provided as inputs, and the combined belief function is monotonic, then overall interpretations are guaranteed to be generated in best-first order. Applications of lazy evaluation to post code dictionary and address recognition problems are outlined as illustrations of the algorithm.