Glossary · Agentic AI audit and record substrate
The terms regulators and auditors use. And the ones we had to coin.
Industry terms are defined as the field defines them. Where we could not find a standard word for a property that matters, we coined one, and we label it as ours rather than dressing it up as industry consensus.
What is an agent audit trail?
The durable, attributable record of what an AI agent did: which actions it took, on what data, under which policy, in what sequence, and with what outcome.
What is audit by construction?
The discipline of producing the audit record as a property of the action itself, not as a downstream reconstruction from logs.
What is a tamper-evident audit log?
A record whose integrity an auditor can verify from any point back to the start.
What is agentic AI accountability?
The ability to answer, for any decision an AI agent influenced, which agent acted, on whose authority, against which policy, with what data, and with what outcome.
What is non-human identity?
The credentials and machine identities under which agents, services, and automation act.
What is a harness autonomy tier?
The level of independence an AI agent has inside the harness that orchestrates it, from operator-approved actions to agents acting on their own authority inside policy.
What is audit-of-action?
The substrate-produced, attributed, tamper-evident record of what an AI agent did.
What is a hash-chained external anchor?
A commitment of a hash chain checkpoint to a separate append-only substrate outside the chain, making wholesale chain replacement detectable.
What is recompute at the anchored recent horizon?
The property that a verifier can independently walk a bounded recent window of a hash-chained record against its external anchor and confirm each entry.
What is verifier separated from emitter?
The architectural property that the process running the recompute is distinct from the emitter that wrote the chain, so the emitter cannot influence the verify.