Safe mode
When the system is uncertain, it stops and asks for the missing pieces.
How we're different
- Uncertainty routes to exceptions, not fabricated output.
- Drafts become checklists when evidence is missing.
- Approvals remain enforced even in demo mode.
Trust rule: AI drafts, humans decide for sensitive outcomes.
What triggers safe mode
- Schema validation fails for an LLM-drafted payload
- Deterministic checks show low confidence / ambiguity
- Required evidence is missing past an escalation window
What you get instead
- A checklist of missing inputs (evidence + owners)
- A proposed next step and escalation path
- An exception inbox item with context and logs