Preserve scope
Attach tenant, app, agent, and pseudonymous subject context so decisions and reusable routes do not cross the wrong boundary.
// AI agent governance
Put policy between agent intent and action.
Punk evaluates model and tool activity with identity-aware rules. Low-risk work can continue; consequential actions can be blocked or held for a person to review.
Punk provides enforcement and evidence primitives. Your organization remains responsible for its policy design, access controls, and compliance decisions.
// controls that travel with the request
Know who, what, and how risky.
Evaluate behavior
Apply MeshGuard-compatible rules to model requests and declared tool actions within the runtime path.
Pause consequences
Hold selected actions for human review instead of treating every agent choice as equally safe.
// conservative defaults
Unknown actions should not look harmless.
Punk uses five side-effect levels, from read-only work through irreversible external consequences. Undeclared tools default conservatively to a user-visible write posture.
| Action posture | Example | Possible handling |
|---|---|---|
| Read-only | Retrieve approved knowledge | Allow and record, subject to policy |
| Reversible write | Update a draft record | Allow, constrain, or review |
| User-visible write | Send a message | Often require explicit policy or approval |
| External commitment | Place an order | Hold or block by default policy |
| Irreversible | Delete production data | Strongest restriction |
// adaptive, not unbounded
Learning never outranks policy.
A reusable route must remain inside the same tenant and safety dimensions, pass its evidence gate, and be allowed by current policy. Ineligible optimized paths return to the configured live provider.
What teams gain
- One replayable record of requests, tools, policy, and routes
- Human review for selected consequential actions
- Route explanations attached to served responses
- Policy-aware replay and shadow testing
Important boundaries
- Auditability is not a compliance certification
- Tool declarations must be accurate
- Policies require testing and ownership
- Fallback behavior must match your risk posture
Observe policy before enforcing it.
Begin with visibility, validate identity and tool declarations, then choose which decisions should block, continue, or wait for approval.