Veritrace

Veritrace was built around a practical question: if an LLM can draft useful rules from text, how canthose rules enter a runtime system without becoming hidden, unreviewed behavior? The answer inthis prototype is a gated pipeline. The LLM drafts, the human reviews, the backend promotes andvalidates, and Jason performs the runtime reasoning.The review console supports two small demo areas. The Cardiac path is the main authoring demo:it has a prepared CMR Mass Score source, claim, and candidate rule for the fast walkthrough, andit can also be used through the live LLM pipeline. GDPR is stronger on the runtime side: it hasapproved rules, approved plan artifacts, benchmark cases, and expected traces for compliancescenarios such as missing legal basis, special-category processing, and late breach notification.At runtime, Veritrace returns a trace with the decision, risk, activated rules, evidence, missing data,source identifiers, human-review flags, and planner-generated next steps. The LLM can be usedagain at the end to turn this trace into a readable explanation, but it receives only the trace andselected source snippets. The trace remains the source of truth.The result is not meant to be a production compliance or medical tool. It is an exam prototypeshowing how LLM-assisted authoring, human review, symbolic execution, and trace inspection canbe connected in one reproducible local workflow.