Verifiable Execution
Can a verifiable execution record prove that an agent never called a specific external API during a run?
The record proves only which steps were executed, not which were skipped. An auditor can conclude that a given API was not called because no signed step record references it, but they must trust the record’s completeness. Catalyst’s record lacks explicit ‘non-invocation’ proofs; absence of evidence is evidence of absence only if the record is known to be complete and unaltered. This limitation means auditors should not rely solely on the record for negative assertions about external interactions.
Was this article helpful?
Your feedback helps improve Diagrid's FAQ experience.
Keep reading
More Diagrid FAQ articles
- Verifiable Execution
What is a verifiable execution record in Catalyst?
A signed, immutable sequence of agent steps enabling auditors to prove run integrity through independent replay. It does not validate decision correctness.
- Verifiable Execution
How does Catalyst ensure that step records are tamper-evident?
Catalyst uses a hash chain and digital signatures per step, so any alteration breaks the chain and becomes detectable through signature verification.
- Verifiable Execution
Can an auditor independently replay a Catalyst agent run to verify its outcome?
Auditors can replay runs using signed step records and initial state to verify output consistency, though non-determinism may limit exact reproduction.