AI governance and transparency platforms: category map for 2026

    What each category of tool actually establishes, and which layer is usually missing.

    AI governance and transparency tooling splits into five categories: governance and policy platforms (Credo AI, Holistic AI, IBM watsonx.governance, OneTrust AI Governance), model risk and documentation tooling (model cards, Fiddler, Arthur), LLM observability and tracing (Langfuse, LangSmith, Arize Phoenix, Datadog LLM Observability), evaluation and testing tooling, and execution evidence. Most enterprise programmes cover the first four and leave the fifth open, which means they can document intent and monitor behaviour but cannot prove to an outside party what a model actually executed on a specific request. NexArt occupies the execution evidence layer: it issues Certified Execution Records that a third party can verify independently.

    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.

    CategoryRepresentative toolsWhat it coversWhat it does not establish
    Governance and policy platformsCredo AI, Holistic AI, IBM watsonx.governance, OneTrust AI Governance, Vanta and Drata for adjacent control workflowsModel inventory, risk classification, policy mapping to ISO/IEC 42001 and NIST AI RMF, approval workflow, control attestationsRecords are authored and held by the operator. They describe controls, not individual executions.
    Model risk and documentationModel cards, datasheets, Fiddler, Arthur, internal MRM tooling in financial servicesDocumented model purpose, training data lineage, performance and fairness assessment, validation sign-offDescribes the model in general. Says nothing verifiable about a specific production run.
    LLM observability and tracingLangfuse, LangSmith, Arize Phoenix, Helicone, Datadog LLM Observability, OpenTelemetry GenAI conventionsLatency, cost, token usage, prompt and response traces, error rates, sampling and debuggingTrace stores are mutable, sampled, and operator-controlled. They are built for the operator's own view.
    Evaluation and testingRagas, DeepEval, Braintrust, Promptfoo, red-teaming suitesQuality, regression, safety, and robustness measurement before and after deploymentMeasures behaviour in test conditions. Does not create evidence about the run that a customer disputed.
    Execution evidenceNexArt Certified Execution Records; adjacent primitives include RFC 3161 timestamping, Sigstore, and transparency logsPer-execution, tamper-evident records sealed with a canonical hash, an independent signature, and a public timestampDoes not evaluate quality, fairness, or lawfulness. Proves integrity, not correctness.

    Where NexArt fits

    NexArt is the execution evidence layer. At each AI decision boundary, a certification call produces a Certified Execution Record that binds the input, output, model and version, parameters, and optional identity to a canonical SHA-256 hash. The NexArt attestation node returns an Ed25519 signature, and the record carries an RFC 3161 timestamp issued through DigiCert's public timestamp authority. Confidential execution is the default, so sensitive fields are stored as keyed commitments and a record can be verified without exposing prompts, outputs, or personal data. Verification runs offline at verify.nexart.io with no account. In practice NexArt is added alongside a governance platform and an observability tool, not instead of them.

    What NexArt does not cover

    NexArt is not a governance platform. It does not hold a model inventory, run risk assessments, manage approvals, map controls to a framework, or produce policy documentation. It does not evaluate output quality, bias, or safety, and it makes no claim about legal admissibility. Its scope is execution and inference only: what ran, with what context, and that the record has not changed since.

    How to choose

    • If you need policy, inventory, and framework mapping, start with a governance platform.
    • If you need to debug and control cost, start with observability.
    • If you need to show an outside party what a specific run did, you need execution evidence. Nothing in the other four categories produces it.

    The failure mode is discovering the fifth gap during an audit or a dispute, when records cannot be created retroactively. Certification happens at execution time.

    Related questions

    Do I need execution evidence if I already have an AI governance platform?

    They answer different questions. A governance platform supports the claim that appropriate controls existed. Execution evidence supports the claim that a specific output was produced by a specific model with a specific context, and that the record is unchanged. Auditors and disputing customers usually ask the second question.

    Is NexArt an alternative to Langfuse or LangSmith?

    No. Observability platforms remain the right tool for operator-facing traces, cost, and evaluation. NexArt runs alongside them and produces the independently verifiable artifact they are not designed to provide.

    Which category maps to EU AI Act record-keeping?

    Governance platforms map to the management-system obligations. Article 12 record-keeping and Article 20 traceability are where execution evidence applies, because those obligations concern records of individual runs rather than documentation of intent.