Goal integrity verification throughout execution

Control objective

Goal integrity verification protects an agent's objective from being tampered with mid-execution. Once a task's goal is set, a long-running agent plans and acts over many steps, and an attacker or a poisoned input could try to silently rewrite that goal - for instance redirecting a payment-reconciliation agent to a fraudulent destination. This control cryptographically pins the goal: it is hashed when authorised, and that hash is re-verified at every planning step so any unauthorised change is detected immediately. To implement, compute and store a cryptographic hash of the approved goal at task start, recompute and compare it before each planning or tool-use step, reject any goal edit that does not originate from a trusted, authenticated source, and log every verification and every rejected edit as tamper-evident evidence. No statistical formula applies; verification is hash-equality. The threshold demands 100% goal-hash verification at each planning step and zero accepted goal edits from an untrusted source. Any hash mismatch or any accepted untrusted edit is a hard breach that must halt the agent immediately and trigger incident response, since it indicates the agent may be executing a hijacked objective.