Browse docsshow

Examples

Five realistic verification patterns. Each one is the same call shape — just different prompts, contexts, and verdict sets.

1. Refund classification

A support agent wants to route customer messages. Refund requests need special handling, but customers rarely use the literal word "refund".

python

tbv.verify(
    prompt="Is this customer message a refund request?",
    context={"message": customer_text},
    allowed_verdicts=["yes", "no"],
    confidence_threshold=0.85,
    source="support/refund-classifier",
)

2. PII leak detection

Before showing an LLM response to an end-user, verify it doesn't contain unmasked personal data.

python

result = tbv.verify(
    prompt="Does this output contain unmasked personally identifiable information?",
    context={"output": llm_response},
    allowed_verdicts=["clean", "leaked", "borderline"],
    confidence_threshold=0.9,
    source="agent/pii-guard",
)
if result["final_verdict"] != "clean":
    return redact(llm_response)

3. Payments fraud screen

Run an LLM-based second-line review on transactions your primary fraud model flagged. Borderline cases route to a human in your fraud-ops team.

python

tbv.verify(
    prompt="Should this transaction be allowed, blocked, or sent for manual review?",
    context={
        "transaction": {
            "card_last4": "8821",
            "amount_usd": 4950,
            "payee": "new (no prior transactions)",
            "location": "Lagos, Nigeria",
            "time_since_login_minutes": 3,
        },
        "cardholder_baseline": {
            "typical_amount_usd": "50-500",
            "typical_payees": "known/recurring",
        },
    },
    allowed_verdicts=["allow", "block", "review"],
    confidence_threshold=0.9,
    source="payments/fraud-secondline",
)

4. Tool-call validation

Before executing a tool an agent picked, verify it's the right tool with valid arguments. Catches the "agent confused refund with cancel" class of bug.

python

tbv.verify(
    prompt="Did the agent invoke the correct tool with valid arguments for the user's request?",
    context={
        "user_request": user_message,
        "tool_called": chosen_tool,
        "tool_args": chosen_args,
        "tools_available": list(available_tools.keys()),
    },
    allowed_verdicts=["correct", "wrong_tool", "invalid_args"],
    confidence_threshold=0.9,
    source="agent/tool-call-validator",
)

5. Decorator / wrapper pattern

Wrap your existing classifier so verification runs automatically. The wrapped function is your existing primary; the decorator forwards its return value AND submits a verification.

python

from tobeverified import verified

@verified(
    allowed_verdicts=["safe", "unsafe", "needs_review"],
    confidence_threshold=0.9,
    source="my-app/comment-mod",
)
def moderate(text: str) -> dict:
    # Build the verification payload
    return {
        "prompt": "Is this comment safe to publish?",
        "context": {"text": text},
    }

result = moderate("user-supplied content")
# result.verdict, result.tier, result.output