Evidence / FIELD NOTES
What does a verifiable AI agent audit trail actually prove?
Separate integrity, attribution, completeness and real-world outcomes when evaluating signed agent records and evidence exports.
A valid signature can support integrity and attribution. It cannot by itself establish that a recorded claim is true, complete or authorized.
“The evidence verifies” is an incomplete conclusion. A reviewer still needs to know what was checked, which key was trusted, which records were included and what the result says about the underlying action. A green verification result is useful only when its meaning is narrow enough to explain.
NIST's Digital Signature Standard describes signatures as a way to detect unauthorized data modifications and authenticate the signatory. That is an important capability, but it is different from independently observing the event described in the signed data. FIPS 186-5 provides the cryptographic foundation for this distinction.
This article proposes a practical review framework. It is not a statement that any particular export format, implementation or product meets every property discussed below.
Start with the exact assertion
Imagine a hypothetical agent record that says an invoice was approved at 10:14. A signature over that record can help establish that the signed content has not changed and that the signature matches a particular key. It does not establish that the invoice system accepted the approval, that the key belonged to the right person, or that the person had authority over this invoice.
Those are separate assertions. The reviewer should be able to identify which evidence supports each one. A trusted key register might establish who controlled the signing key during the relevant period. A policy record might establish permitted scope. A destination event might establish that the invoice system recorded a matching transition.
The design mistake is asking one cryptographic result to stand in for all of them.
Review four dimensions separately
Integrity: Do the provided files match the signed material? Are the expected records and their relationships intact under the format's rules? A failed check should name the affected object and the reason, rather than merely display a red badge.
Attribution: What connects the signing key or credential to an accountable identity? Was that connection valid at the time? “Key verified” and “employee verified” should not be interchangeable labels.
Scope and completeness: Which source, tenant, interval and filters define the export? What could have been excluded before signing? A perfectly intact export can contain an incomplete selection. If a collector was offline, the export should carry that limitation rather than imply that no activity occurred.
Outcome: What independent record supports the claimed result? For a business mutation, this may be the destination's audit event or a controlled readback. Its reliability still depends on access, retention, interpretation and the identity of the source.
These dimensions are our suggested review structure. They are intended to prevent stronger conclusions than the evidence supports.
Preserve the difference between missing and contradictory evidence
Suppose the signed agent record says “approved,” while the destination says “rejected.” That is a contradiction to investigate. Suppose instead that the destination connector was unavailable. That is a collection gap. Treating both as “unverified” without further explanation discards the information needed to respond sensibly.
A useful export can carry the original record, the destination observation, the matching rule and the resulting assessment. If a later observation changes the assessment, preserve the earlier observation and attach the update. Reviewers should be able to see why the interpretation changed.
Give the reviewer a reproducible path
We recommend documenting the verification procedure alongside the evidence format. State which files are inputs, how keys are selected, how relationships are checked and what each failure means. Provide a known-good sample and a deliberately modified sample so a reviewer can check that the verifier detects a relevant change.
Also test the boundaries. Remove a file. Supply an unknown key. Alter a record without altering its signature. Change the claimed scope. The expected result should be explicit for each case. These tests do not prove every property of the implementation, but they make the intended contract inspectable.
Keep the verification tool's own version visible. A result without a tool version, input identity or stated scope is hard to reproduce months later.
Write conclusions that survive scrutiny
Prefer a conclusion such as: “The supplied files match this signed manifest, using this trusted key. The export covers the stated source and interval. Three actions lack destination observations.” That gives a reviewer something concrete to accept, challenge or extend.
Avoid saying “all activity is proven” merely because the provided bundle passes integrity checks. The absence of a record cannot establish the absence of an event unless the collection and reconciliation contract support that inference.
The strongest evidence systems make their limits easy to find. They allow a reviewer to distinguish an unchanged claim from a confirmed result, and a well-defined scope from an assumption of completeness. That precision is what makes verification useful beyond the team that generated the export.
Sources and further reading
Primary sources checked on 16 November 2026. Examples in this article are illustrative unless stated otherwise.