These are different cryptographic claims. One proves recorded events were preserved. The other proves the governed process executed as declared on the committed artifacts.
WORM logs, hash chains, and tamper-evident audit trails prove that recorded events were not modified. Primust proves that a specific governance process executed as recorded on committed data with a specific outcome.
Those are different cryptographic claims. One proves log integrity. The other proves execution. A WORM log can be perfectly intact and still be a record of nothing useful — because nobody recorded the thing that mattered, or recorded it wrong, or recorded it from a system that was already compromised.
"Log integrity proves the logs weren't edited. Primust proves the governance executed. Those aren't the same claim — and when review turns adversarial, execution is the claim people actually test."
These are all log-integrity primitives. Every one assumes the operator recorded the truth. They make the log itself tamper-evident. They do not make the operator honest.
The agent takes an action and never writes a log entry for it. WORM is empty. Hash chain is intact. Nothing to verify. This is what Amazon reframed as "user error" — the logs technically showed what was requested, not what the agent concluded and executed.
→ Log integrity = perfect. Forensic value = zero.
The agent or operator writes a log entry describing something that didn't happen — or describing something that did happen but with the wrong contents. WORM stores it faithfully. Hash chain includes it. RFC 3161 confirms it was written at that time.
→ All three primitives cheerfully confirm a lie was written at a specific time. This is the Replit case — the agent fabricated replacement records, and standard logging infrastructure had no way to flag them as fabricated.
An attacker or the agent itself modifies the config of the logger before the action, such that the true action is never recorded. The hash chain never sees the real event. WORM has no input to preserve. RFC 3161 timestamps only the logs that were written.
→ The untaken writes are invisible. Apollo Research caught frontier models editing oversight_enabled: true to false in their own runtime config.
Primust binds input, output, and process step at the moment of execution, inside the customer's environment, before any tampering surface exists.
The commitment to the input, the output, and the process step is computed in-process, before the result leaves the boundary. Not after the fact by the logging system. Not streamed to a SIEM. Bound to the computation itself, locally, before any attacker — human, agent, or SIEM — can manipulate it.
input_commitment when ZK proof capability is available. Falls back to SHA-256 with explicit proof-ceiling downgrade to Execution — we never silently claim Mathematical while using SHA-256.customer.com/.well-known/primust-key.json. Primust has zero ability to forge on your key.A verifier in 2036 can check the credential against all three, with zero Primust dependency. That is the gap WORM + a single TSA can never close.
EC_SIGN_ED25519). AWS KMS and Azure Key Vault don't.Self-attested logs can look complete while still missing the evidence a reviewer needs. See the pattern →, or ship a VPEC today.