Agent Decision Policy Binder

Check the relationship between a normalized decision event and a redacted policy definition.

Scenario and prerequisites

• - Use after context normalization and before any real action or export.
• Provide a policy snapshot with version, allowed_actions, required_permissions, approval_required and owner.
• This is a deterministic evidence check, not an authorization service.

Input and output

Input is the first skill output plus a policy JSON file. Output includes AUTHORIZED or HOLD_FOR_HUMAN_REVIEW, the binding hash, matched permissions and reasons.

Installation and execution

1. 1. Keep the skill directory and policy file in a controlled local workspace.
2. Run python scripts/bind_agent_policy.py --event ../agent-decision-context-normalizer/examples/output.json --policy examples/policy.json --output examples/output.json.
3. Pass only an AUTHORIZED or explicitly reviewed hold record to the chain verifier.

Recovery and boundaries

Version mismatch, disallowed action, missing permission or required approval produces a hold. A hold is not an allow and must not be converted by editing the output.

The script does not call IAM, execute tools or grant permissions. It evaluates only the supplied redacted policy snapshot.

Acceptance

The sample policy authorizes the sample action and emits AUTHORIZED. Removing the listed permission or changing the policy version must emit HOLD_FOR_HUMAN_REVIEW.