Evidence needs a scope.

A record can be authentic and still fail to answer the question a reviewer cares about.

Axiom is an exploratory line of work about capturing evidence from ML and agent workflows. The central question is practical: what must remain after a run so that someone else can interpret it?

Identity comes first

A useful record connects the execution request, inputs, model artifacts, runtime settings, outputs, and errors. A checksum helps identify content. A timestamp helps place an event. Neither substitutes for the relationship between them.

Missing context is a result

If an input was never recorded, a dependency cannot be recovered, or the execution context is unknown, the record should say so. An explicit gap is more useful than a confident reconstruction built on assumptions.

• Which inputs and artifacts does the record identify?
• What settings and runtime context were captured?
• Which outputs and failures belong to this attempt?
• What can a reviewer not reconstruct from what remains?

Signatures have a specific job

Schema validation, hashes, and signatures help check the recorded content. They do not guarantee deterministic replay, establish the truth of an execution, or certify regulatory compliance.