Reasoning trace logging for consequential decisions

Control objective

Reasoning trace logging for consequential decisions checks that whenever an agentic AI makes a high-impact call — approving or declining a facility, flagging a transaction, escalating a case — the system captures a complete, tamper-proof record of how it got there: the inputs considered, the retrieved evidence, the intermediate steps, and the final output, so the decision can later be reconstructed exactly. This matters because opaque automated decisions are indefensible to auditors, regulators, and affected customers; without a replayable trace a bank cannot explain or contest an adverse outcome. No metric or formula is specified for this control, so it is governed directly by its threshold rather than a computed score. Implement it by hooking the agent's decision pipeline so that every above-threshold decision emits a structured trace, writing those traces to an append-only or immutable store, linking each to the decision id, and providing tooling to replay them. The threshold requires that 100% of above-threshold decisions have an immutable reasoning trace that is replayable for audit; any consequential decision found without a recoverable, replayable trace is a breach that must trigger investigation and a fix to the logging path, since it represents an unauditable automated action.