Operate within scope
A request may proceed automatically only when it fits rules and permissions established in advance.
Governed AI · authority before execution
Decide what AI may do before it acts.
SentientROUTER evaluates a request against institutional policy and routes it to permitted automation, a draft for review, or a human decision. The authority boundary remains explicit at every step.
The authority path
01
RequestAn institution defines the action and its context.
02
Policy checkConfigured rules set the permitted route.
S
Permitted automationOnly within approved scope.
D
Draft for reviewNo autonomous final action.
H
Human approvalExecution waits for an authorised person.
Decision model
Three routes, one clear boundary.
The S, D and H classes express what an AI workflow is allowed to do next. Classification is a governance decision, not a score of model quality.
Prepare, then stop
The system may assemble an answer or recommendation. A person reviews it before any consequential action.
Require a person
A designated human authority must approve or decline the action. An unapproved request cannot simply move to a lower class.
How it works
From request to accountable action.
The sequence describes the intended control architecture. The controls and integrations used in a particular institution require evaluation and configuration.
Capture the request
Identify the proposed action, source, relevant data and institutional context.
Apply policy
Check the request against the current version of the institution's rules and delegated authority.
Choose a route
Return S, D or H with the reason and the governing rule. Uncertain or incomplete context escalates.
Gate execution
Permit a bounded action, prepare a draft, or wait for a person with the required authority.
Record the outcome
Keep the classification, approvals and execution context available for institutional review.
Review and adjust
Assess incidents, policy changes and behaviour before expanding any permitted autonomy.
Architecture
Separate policy from model output.
The control plane is designed to sit before AI driven actions and make authority visible to the people who operate and review the system.
Rules and versions
A policy version and its scope can be identified when a request is classified.
Human decision points
High impact actions can be held until the appropriate person makes a recorded decision.
Conservative escalation
Divergence, missing context or governance concerns can narrow automation rather than silently expanding it.
Reviewable events
The classification and its context can be passed to an evidence layer for later checking.
Related technology
Control and proof have different roles.
SentientROUTER describes whether an AI action may proceed. ProvableCORE describes how evidence of a governed event may be bound, retained and verified.
The control plane
Routes a request under institutional rules and human authority.
The proof layer
Binds an event to its context, policy, approvals and integrity evidence for later verification.
The sites describe the intended architecture of the product family. They do not assert a live production integration or a certified deployment.
Institutional context
Designed for institutional model governance.
Human approval, traceability and effective review are relevant design considerations for institutions in the Americas. Each deployment requires its own legal and risk assessment.
Defined responsibility
The institution determines which roles may approve a consequential action and under which policy.
Reviewable history
A reviewer should be able to see the policy, classification and human decision associated with an event.
No certification, regulatory approval or legal conformity is asserted by this description.
Illustrative example
A credit request held for review.
This schematic example shows a possible policy outcome. It is not a live API response, production decision or configured credit rule.
request: agricultural_credit_approval policy: institution_policy_v1 route: H autonomous_action_permitted: false human_approval_required: true evidence_state: pending_verification
A real institution would define the policy, threshold, authorised role and evidence requirements before using the workflow.
Questions
What institutions should verify.
Does this replace an institution's approval process?
No. It is intended to encode the permitted route and preserve a clear handoff to authorised people. Approval rules remain the institution's responsibility.
Can model confidence increase autonomy?
Confidence alone cannot expand delegated authority. A change in autonomy requires an explicit policy and approval process.
Is the full platform already qualified for my organisation?
No general claim of qualification is made. Technical controls, security, legal obligations and integrations must be assessed for each use case.
How does ProvableCORE fit in?
It is the related proof layer. Whether and how the two components integrate in a specific deployment must be established in that deployment.
Contact
Discuss a controlled evaluation.
Tell us about the intended use case and the authority rules you need to enforce. We can discuss the architecture and the evidence needed for an evaluation.