Company policy caps agent-issued refunds at $500; anything larger must go to a human. The system prompt states this rule clearly, yet an audit finds a handful of process_refund calls for $700-$1,400 each month, usually in long conversations with persuasive customers. What is the most effective fix?

Answer: B — Add a hook that intercepts outgoing process_refund calls, blocks any whose amount exceeds $500, and redirects the case to escalate_to_human with the requested amount attached.

A hard business rule with financial consequences needs a deterministic guarantee, and a tool-call interception hook gives that: the oversized call never runs, and the case is sent on to escalation (task 1.5). Prompt wording and few-shot examples (a) are still probabilistic, and the audit already shows a non-zero failure rate for prompt-only enforcement. Self-reported confidence (c) is poorly calibrated and is still the model policing itself. An iteration limit (d) is an arbitrary cap that doesn't address the rule, cuts legitimate long conversations short, and a short conversation can still produce an oversized refund.