User intent, role context and explicit decision points.
CONTROLLED LAYER08
Reference Architecture
Intelligence belongs inside a controlled decision path.
Our reference architecture separates reasoning from authority. Models may recommend, plan and interpret; policy and accountable owners determine what the system is permitted to do.
Workflow state, planning, routing, memory boundaries and recovery.
CONTROLLED LAYERSelected models, structured reasoning, retrieval and confidence signals.
CONTROLLED LAYERPolicy engine, tool permissions, approval gates and transaction limits.
CONTROLLED LAYERSources, citations, evaluation traces, decision logs and audit events.
CONTROLLED LAYERObservability, incident response, rollback, cost and performance controls.
CONTROLLED LAYERCONTROL PATH
From request to accountable action.
Authenticate identity and business context
Retrieve only permitted, relevant evidence
Generate a structured proposal with uncertainty
Evaluate policy, risk and tool permission
Request human authority when threshold requires
Execute, verify outcome and preserve an audit event
FAILURE CONTRACT
What happens when the system does not know?
- Refuse unsupported conclusions rather than manufacture certainty.
- Degrade to a safer workflow when a dependency or model fails.
- Escalate consequential ambiguity to a named human role.
- Prevent partial tool execution from becoming silent state corruption.
- Record exceptions so failure improves the next release.