How Sentinel SCA Works
Sentinel SCA governs autonomous actions before they execute. It separates decision, enforcement, receipt verification, and audit integrity into a deterministic governance flow.
The short version
Agent proposes action ↓ Sentinel verifies identity and capability ↓ Sentinel evaluates schema, risk, and policy ↓ Decision: ADMIT / REVIEW / DENY ↓ Execution boundary enforces the decision ↓ Boundary receipt returns to Sentinel ↓ Audit chain and replay evidence are preserved
1. Agent proposal
Sentinel starts when an AI agent, service, device, or autonomous system proposes an action. The proposal may involve software, infrastructure, physical devices, APIs, or workflows.
2. Identity and capability verification
Sentinel first asks who is making the request and what that actor is allowed to do. Identity alone is not enough. A valid agent must also have the correct capability for the requested action.
3. Schema, risk, and policy evaluation
Sentinel validates the structure of the request, evaluates its risk level, and applies deterministic governance policy. This produces a governance decision rather than a simple log entry.
ADMIT — the action is allowed to proceed.
REVIEW — the action is halted until human approval.
DENY — the action is blocked.
4. Execution boundary
Sentinel is designed around the execution boundary: the point where a proposed action becomes real. This may be a device action, infrastructure command, external API call, financial operation, or production workflow.
5. Boundary receipts
When an external boundary receives or enforces a governance decision, it can return a receipt to Sentinel. Receipts show whether execution was halted, admitted, executed, rejected, or expired.
6. Audit chain and replay
Sentinel preserves governance evidence in audit and integrity records. This supports replay, forensic review, and verification of why an action was allowed, denied, or routed to human review.
7. Fail-closed enforcement
When Sentinel cannot verify identity, capability, policy, timing, evidence, or execution status, the safe outcome is non-execution. This is fail-closed governance.
Example: governed fan activation
In an autonomous agricultural environment, an edge node may propose fan activation when temperature rises. Sentinel verifies the node, checks its environmental-control capability, evaluates policy, and then admits, denies, or routes the action for review.
Temperature high ↓ Agent proposes activate_fan ↓ Sentinel evaluates ↓ Decision: ADMIT ↓ Execution boundary receives command ↓ Receipt recorded ↓ Audit chain updated
Why this matters
AI monitoring explains what happened after execution. Sentinel governance determines whether execution should happen at all.