The thesis
The problem is not simply that a model can be wrong. It is that an application can treat an unsupported output as an instruction, a fact, or a decision. I developed LSAS to make that transition explicit in regulated healthcare and medtech settings. The specification describes five stages: classify intent and risk, ground claims in authoritative sources, validate the output against relevant constraints, escalate unresolved cases to human review, and improve versioned policies through regression testing. Each stage has a distinct responsibility. A content filter cannot establish that a claim is supported; a citation cannot establish that an action is authorized.
Why it matters
What exists today is a public specification repository with a technical whitepaper, schemas, examples, and a citable Zenodo record. That is a concrete architectural artifact, not a claim of clinical validation or universal safety. LSAS makes the argument behind “Stop Trying to Make AI Safe” inspectable: engineer the release boundary, make uncertainty visible, and retain evidence of why an output was allowed to proceed. Bounded autonomy and auditability support that approach in agent systems. Teams must still implement the contracts, test their failure paths, and provide the people and operating procedures that escalation requires.