I chair the audit committee. When the external auditor, a regulator, or a plaintiff asks how this board used AI in its decisions, what evidence can I actually produce — and will it stand up?

The full executive answer

You can produce a complete, tamper-evident reconstruction, and the tamper-evidence is the part to examine closely because it is what separates an audit trail from a log file. Every material event on the platform — generation, edit, approval, access, export — is written to an append-only audit log where each record is cryptographically chained to its predecessor by hash: alter or delete any historical record and the chain visibly breaks. The log is written through a privileged service path the application cannot mutate, and "append-only" is enforced in the database, not promised in a policy. For any decision, you can therefore produce: the inputs, the generated analysis with every number’s provenance to source, the cross-model consensus results including disagreements and any third-model adjudication, the named human approvals with timestamps, and every subsequent access — the full chain of custody an evidentiary standard asks for.

On "will it stand up" — the honest, two-part answer. Technically, hash-chained append-only records are the same integrity pattern regulators already accept in financial audit logging under SOX-era expectations and in SOC 2’s processing-integrity criteria; nothing about the mechanism is novel to an assurance professional, and your auditor can verify the chain independently. What we do not yet have is precedent — no court or regulator has yet tested this specific record in anger, because we are pre-launch; I will not pretend otherwise. What I can offer your committee today is inspectability: under NDA, your audit function can review the audit-log design, the erasure code, and the CI test that fails our build if any customer-data table lacks isolation enforcement. We built for the day someone hostile reads the log; I would rather your auditor be the first hostile reader than a plaintiff’s expert.

One more property an audit chair will value: the log also captures what the AI was not allowed to do — guardrail blocks, provenance failures, rejected generations. Evidence of controls operating, not merely existing, is precisely the distinction your external auditor draws between design effectiveness and operating effectiveness, and the platform generates the operating evidence continuously as a by-product of running.