AI audit trail requirements in banking, and where logs stop being enough

Supervisory expectations for AI in banking are converging on a familiar shape: know what the system does, govern it, document it, and be able to reconstruct specific decisions on request. Institutions generally handle the first three. The fourth is where examinations get uncomfortable.

What a reconstructable decision requires

• The model and version in force at the time.
• The inputs, or a cryptographic commitment to them where the inputs are sensitive.
• The output, including confidence or score where applicable.
• Human review, override, or documented absence of either.
• A timestamp that is credible to someone outside the institution.

The first four are ordinary engineering. The fifth is the hard one, and it is the one that carries evidentiary weight.

The self-attestation gap

An institution's logs are produced, stored, and maintained by the institution. In a dispute about whether a decision happened as described, that provenance is exactly the weakness opposing counsel will press. Tamper-evidence that depends on the log owner's controls is not tamper-evidence to a third party; it is a policy assurance.

Independent verifiability is what closes the gap. If the record's timestamp and integrity are witnessed by a party with no interest in the outcome, the institution's own controls stop being load-bearing.

What this looks like operationally

Rubric attests decision records at the moment they occur: the record is signed with a post-quantum key the application does not hold, and the signature is anchored to the Hedera Consensus Service, a public ledger operated by an independent council. Anyone can later verify that a record existed at a given time and has not changed, using public infrastructure.

Where institutions usually start

Attesting every inference is rarely the right first step. The high-value scope is the set of decisions that are expensive to be unable to prove:

• Adverse actions and denials.
• Autonomous approvals above a materiality threshold.
• Model version changes and configuration edits.
• Incidents, overrides, and escalations.

Evidence quality matters more than evidence volume. A small set of independently verifiable records is worth more in an examination than terabytes of logs nobody can authenticate.

Standards posture

Evidence packages produced by Rubric carry C2PA Content Credentials, the same content provenance standard adopted across the major technology platforms, on top of the post-quantum attestation layer.