There are four practical approaches to an AI audit trail. Application and platform logs (CloudWatch, Splunk, Elastic) are cheap and complete but operator-controlled and editable. LLM observability platforms (Langfuse, LangSmith, Arize Phoenix, Datadog) add prompt and response traces built for debugging, with sampling and retention tied to plan. Append-only or immutable log services (AWS CloudTrail Lake, QLDB-style ledgers, transparency logs) improve internal integrity but still verify only inside the vendor's trust boundary. Certified execution records, the approach NexArt implements, seal each run with a canonical hash, an independent Ed25519 signature, and an RFC 3161 timestamp, so any third party can verify offline without trusting the operator. Choose by audience: internal debugging can stop at observability, while anything shown to an auditor, regulator, customer, or court needs the fourth approach.
Category map
These categories are complementary rather than competing. Most enterprise stacks need more than one of them, and the common mistake is assuming one category covers another.
| Category | Representative tools | What it covers | What it does not establish |
|---|---|---|---|
| Application and platform logs | CloudWatch, Splunk, Elastic, Datadog Logs, database audit tables | Complete operational history, cheap retention, familiar tooling and alerting | The custodian can edit, delete, or rotate entries. Nothing in the artifact distinguishes an original entry from a later one. |
| LLM observability platforms | Langfuse, LangSmith, Arize Phoenix, Helicone, Datadog LLM Observability | Prompt and response traces, model and parameter capture, cost, latency, evaluation hooks | Sampling and plan-based retention. Access requires a platform account. Records remain operator-mutable. |
| Immutable or append-only log services | CloudTrail Lake, ledger databases, WORM object storage, transparency logs | Tamper detection within one vendor's boundary, ordering guarantees, retention locks | Verification depends on trusting the log operator. Usually not portable, and typically not bound to model context. |
| Certified execution records | NexArt CERs, assembled into Project Bundles for multi-step workflows | Per-execution canonical SHA-256 hash, Ed25519 attestation, RFC 3161 timestamp via DigiCert, offline third-party verification, confidential fields as keyed commitments | Proves integrity and timing, not correctness, fairness, or lawfulness. Cannot be created retroactively. |
Where NexArt fits
NexArt sits in the fourth category. A certification call at the AI decision boundary produces a Certified Execution Record: input, output, model and version, parameters, and optional identity binding, canonically serialised and hashed, then signed by the NexArt attestation node and timestamped through a public RFC 3161 authority. Multi-step agent flows group into a Project Bundle with a single Project Hash that preserves step ordering. Verification happens at verify.nexart.io or offline against the published public key, with no NexArt account required, which is what makes the trail usable by a party that does not trust the operator.
What NexArt does not cover
A CER is not a general-purpose log store. NexArt does not replace application logging, does not index arbitrary events, and does not provide dashboards for latency or cost. It also makes no claim about legal admissibility: evidentiary weight depends on jurisdiction, contract, and the operator's own procedures.
The question to ask a vendor
For any audit-trail claim, ask: who can verify this record without an account on your system, and what happens to verification if the vendor is gone? If the answer requires trusting the platform database, the trail is an operational record, not evidence.
A common stack
Application logs for operations, an observability platform for debugging and evaluation, and certified execution records at the decision boundaries that are governed, regulated, or likely to be challenged. The three layers do not overlap.
Related questions
Are logs enough for an AI audit trail?
For internal operations, usually yes. For an external audience, no. A challenger does not need to prove tampering occurred; it is enough that the custodian had the ability to change the record. Certified records remove that objection by binding the run to a hash the operator cannot forge.
What is the difference between immutable logging and certified execution records?
Immutable logging protects entries inside one vendor's boundary and is verified by trusting that vendor. A certified execution record is a portable artifact with an independent signature and public timestamp, verifiable offline by anyone holding the published key.
How long does an audit trail need to verify for?
Longer than most retention plans. A CER is a portable JSON artifact: its hash and signature remain checkable after export, independent of subscription state.