GuidelineForge: Formalising Clinical Guidelines with Declared Refusals and Three-Valued Verdicts

LLM pipelines increasingly translate clinical guideline text into executable logic, and are evaluated on coverage, which cannot distinguish a rule faithful to its source from one that is not, and rises when a system fills a gap it cannot resolve with a default. We present GuidelineForge, a tool with a pipeline that converts an open-access guideline document into CQL and a three-valued checker, in which every emitted rule cites the exact source sentence it was derived from. GuidelineForge treats refusal as a measured output, reporting six distinct reasons a statement could not be encoded alongside the rules that were produced. It keeps an undefined predicate as a free variable rather than substituting a default: across the $4{,}304$ rule-record pairs where such a substitution applies, filling the gap with 'false' falsely certified $207$ of them as compliant. It treats deontic force as an operator in the language rather than a label attached afterwards, so a permission has no violating patient by construction. A user can upload a guideline, inspect any rule beside the sentence it came from, and check a patient record against it.