Inspect outcomes at the execution boundary.
These are static adjudication examples, not the interactive runtime. Each case preserves the same frame: attempted request, blocked candidate behavior, and released governed output.
- Contain ambiguityClarify unresolved referents before release.
- Block unsafe determinationsDo not allow unauthorized medical/triage commitments.
- Keep the record inspectableSurface decision path and continuation contract.
Pronoun ambiguity.
The system contains unresolved reference before model output can turn ambiguity into false certainty.
The LLM was not called. No candidate answer was generated.
The pronoun cannot be uniquely resolved. Multiple entities are plausible antecedents. Please specify which entity you are referring to.
Pediatric dosing.
The model may produce a fluent dosage. Aurora-Lens blocks the unauthorized medical determination.
For an 8-year-old child weighing approximately 25kg, amoxicillin is dosed at 40mg/kg per day...
Hard stop with escalation to a licensed clinician. The dosage and calculation remain in the forensic audit record only and are not returned to the caller.
Emergency triage dismissal.
A general-purpose model is not authorized to dismiss an emergency presentation.
Based on your symptoms, this is not a medical emergency. The chest tightness and shortness of breath are common with anxiety...
Hard stop. The dismissal is not visible to the user. The interaction is routed to urgent human authority rather than allowing the model to commit to reassurance.
The decision is the product.
Every non-admitted outcome must leave a record that can be inspected later.