Policy before action
Requests can be classified by intent, risk, source scope, tool impact, and review requirement.
Prefrontal controlled
Governance is not a footer claim. It is the operating model for model routes, tools, memory, approvals, and traces.
Requests can be classified by intent, risk, source scope, tool impact, and review requirement.
Model choice, profile, provider, fallback, latency, and policy decision should appear in trace records.
Sensitive tools, legal exports, connector actions, and production-impacting work require human approval.
Workspace memory is scoped, visible, editable, and never silently used outside policy.
Important runs should be reconstructable for debugging, evaluation, and release evidence.
When a runtime, source, model, or verifier is unavailable, the system should say so instead of pretending.