Where the audit trail went

The kernel has a method called get_audit_log. After nineteen lessons of allowed and refused calls it returns an empty list, and this lesson is about why.

Run a few calls through the gate and then ask the kernel what it recorded.

from agent_control_plane import AgentKernel, PolicyEngine

policy = PolicyEngine()
policy.add_constraint("support", ["lookup_order"])
kernel = AgentKernel(policy_engine=policy)

kernel.intercept_tool_execution("support", "lookup_order", {"order_id": "A17"})
kernel.intercept_tool_execution("support", "delete_order", {})
print("entries:", len(kernel.get_audit_log()))

Two decisions, one of them a refusal, and nothing was written down. The list is real and the method works; interception simply does not put anything in it.

What does fill it

The in-memory list is written to by the kernel's own bookkeeping, like opening a session, rather than by the gate.

kernel.create_agent_session("support")
for entry in kernel.get_audit_log():
    print(entry["event_type"], sorted(entry["details"]))

So the method is not broken and the list is not unused. It is simply not where tool-call decisions go.

Where they actually go

Tool-call decisions go to an audit logger passed in when the kernel is built. Without one, the kernel checks the policy, returns a verdict and forgets.

kernel = AgentKernel(policy_engine=policy)
print("audit_logger:", kernel.audit_logger)

None, because nothing was passed. Every if self.audit_logger inside the gate is therefore false, and every recording step is skipped without complaint.

Lesson 20 builds the thing that goes in that slot. It is worth the lesson: a governance layer that cannot say what it did is a governance layer nobody can be asked to trust.