Policy Routers
Routers are not categories — they are runtime reasoning modules. Each one takes a request context as input and returns a structured verdict through a common contract. For any request, their outputs are adjudicated side by side.
Common Router Contract
Input
Request Context
Prompt, identity, tenant, model, tools, data scope, trust level and environment — the full runtime bundle.
Output
Router Outputs · Side by Side
Evaluated request
Write a PowerShell script to test MFA resilience against our staging identity environment.
SecOps Copilot · j.okafor@acme.com · Staging · isolated sandbox
Cyber Security
cyber · router
Recommended
Confidence
84%
Privacy
privacy · router
Recommended
Confidence
82%
Tool Use
tool_use · router
Recommended
Confidence
87%
Policy Adjudicator
Combines router outputs and resolves conflicts into one accountable outcome.
Risk is supporting evidence. Policy adjudication determines the outcome.
Router Inputs
Cyber Security
DUAL_USE_SECURITY_TESTING
Privacy
SENSITIVE_DATA_PRESENT
Tool Use
SANDBOXED_EXECUTION_ONLY
Policy Precedence
More permissive → more restrictive
The most restrictive content verdict wins.
Final Adjudication
Outcome
Allow with Constraints + Execution Needs Approval
Confidence
84%
Risk (evidence only)
38Moderate
Adjudication Logic
- 1No hard-stop policy fired.
- 2Content precedence took the most restrictive verdict across 2 content router(s): Allow with Constraints (Cyber Security).
- 3Action axis adjudicated independently by the Tool Use router: Allow with Approval.
Module Registry
Governs offensive and defensive security requests. Distinguishes authorised defensive testing from operations that materially enable intrusion, credential theft or disruption of systems the requester does not own.
Subcategories — 12
Evaluation Dimensions
Available Modules — 11