Verdicts

Every governance evaluation returns one of three verdicts. Each verdict tells your application exactly what to do next.

The Three Verdicts

PROCEED

{ "verdict": "proceed" }

All 8 governance laws passed. The action is safe to execute immediately.

Your code should: Execute the action and record the result.

if result.passed:
    execute_action()

---

ESCALATE

{ "verdict": "escalate" }

One or more laws returned a warning, but none failed outright. The action is not clearly dangerous, but it requires human judgment.

Common reasons for escalation:

• Medium-risk action without prior approval
• AI agent approaching the boundary of its autonomy
• Action cost is close to the budget limit
• Missing optional context that would improve evaluation

Your code should: Queue the action for human review. Do not execute automatically.

if result.needs_approval:
    approval_id = queue_for_human_review(result)
    # When approved, re-evaluate with has_approval=True

---

BLOCK

{ "verdict": "block" }

One or more laws failed. The action violates a governance constraint and must not be executed.

Common reasons for blocking:

• High-risk irreversible action without human approval
• Actor does not have authority for this action
• Budget would be exceeded
• Safety concern detected
• Audit trail cannot be maintained

Your code should: Reject the action, log the reason, and alert the operator if appropriate.

if result.blocked:
    log_blocked_action(result.explanation)
    alert_operator(result)

---

Verdict Decision Logic

Condition	Verdict
All 8 laws return pass	proceed
Any law returns warn, none return fail	escalate
Any law returns fail	block

Per-Law Results

Each law evaluation returns one of three results:

Result	Meaning
pass	Law is fully satisfied
warn	Law is conditionally satisfied — human review recommended
fail	Law is violated — action must not proceed

Inspecting Why

The evaluations array in the response shows exactly which laws contributed to the verdict:

result = client.governance.evaluate("data.delete", description="Delete user data")

# Find which laws caused the verdict
for evaluation in result.evaluations:
    if evaluation.result.value != "pass":
        print(f"{evaluation.law_name}: {evaluation.result.value}")
        print(f"  Reason: {evaluation.reason}")

Example output for a blocked action:

Safety Dominance: fail
  Reason: High-risk irreversible action requires human approval
Bounded Autonomy: warn
  Reason: AI agent trust score below threshold for data deletion

Handling Escalations

When an action is escalated, you can resolve it by re-evaluating with additional context:

# Original evaluation — escalated
result = client.governance.evaluate(
    "payment.process",
    description="Process $4,200 vendor payment",
    context={"risk_level": "medium", "is_irreversible": True},
)
# result.verdict == "escalate"

# After human reviews and approves:
result = client.governance.evaluate(
    "payment.process",
    description="Process $4,200 vendor payment",
    context={
        "risk_level": "medium",
        "is_irreversible": True,
        "has_approval": True,  # Human approved
    },
)
# result.verdict == "proceed"

Best Practices

1. Always handle all three verdicts — never assume an action will proceed
2. Log blocked actions — they indicate potential security issues or misconfiguration
3. Track escalation rate — a high escalation rate may indicate your agents need tighter scoping
4. Provide rich context — more context leads to more accurate evaluations
5. Use the explanation — show it to operators so they understand why an action was blocked or escalated