Why AI decisions need tamper-evident evidence
AI systems take consequential actions: approving transactions, modifying infrastructure, accessing sensitive data, controlling physical systems. When something goes wrong — an unauthorized action, a policy violation, a regulatory inquiry — the first question is: what happened, who authorized it, and can you prove it?
Standard application logs cannot answer this question with the required confidence. They can be modified, deleted, or backdated. They lack tamper-evidence. For AI systems operating in regulated environments or taking consequential actions, the audit trail must provide cryptographic assurance that the record has not been altered since creation.
Components of a complete AI audit trail
A complete AI audit trail records, for every consequential action: (1) the proposed action and its full context — what was requested, by whom, with what parameters; (2) the authority evaluation — what authority the actor held, whether it was sufficient, and the delegation chain; (3) the policy evaluation — which policies were evaluated, their resolution, and the effect precedence; (4) the decision — authorized, modified, escrowed, review required, routed, or denied; (5) the evidence considered — context snapshots, risk scores, and external inputs; (6) the execution outcome — what actually happened; and (7) the chain of custody — who accessed the record and when.
Each entry is cryptographically linked to the previous entry via SHA-256 hashing. The trail is append-only: events can be added but not modified or deleted. Database triggers prevent UPDATE and DELETE operations on the trace table.
Deterministic replay
A tamper-evident audit trail enables deterministic replay: reconstructing any decision exactly as it occurred, using the same inputs, policies, and context. This supports incident investigation (what happened and why), regulatory inspection (demonstrating that the system operated as described), and policy testing (evaluating how a policy change would have affected past decisions).
Deterministic replay requires that the audit trail captures not just the decision, but the complete decision context: the policy set in effect, the authority grant, the context snapshot, and the resolution logic. Without the full context, replay is approximate rather than exact.
Audit trails in DecisionHypervisor
DecisionHypervisor maintains append-only, SHA-256 hash-chained decision traces. Each event hashes the previous event's hash, creating a tamper-evident chain. Database triggers block UPDATE and DELETE operations on the trace table. The system supports deterministic replay of any decision, reconstructing the exact policy evaluation, authority check, and context that produced the original outcome.
The trace format is designed for regulatory inspection: a reviewer can follow the chain from any decision back to the originating authority, verify that each hash links correctly, and confirm that no entries have been modified or deleted. This provides the tamper-evidence property that AI audit trail requirements demand.