nexart.io

    Certified AI Evaluation

    Test results should be verifiable, not just reported. When you claim a model performed well, the evidence should be independently confirmable.

    When a team reports that a model achieved a certain accuracy or safety score, the result is only as credible as the process that produced it. Most evaluation pipelines generate self-reported numbers with no mechanism for independent verification. CERs close that gap: every test run becomes a sealed, verifiable artifact that proves exactly what was tested, with what parameters, and what the model returned.

    The evaluation trust gap

    The core problem is not running evaluations. It is defending reported results after the fact. When results are questioned, most teams have no verifiable evidence to point to.

    • Self-reported results. The team that ran the evaluation also reports the numbers. There is no independent verification mechanism.
    • Missing parameters. Temperature, prompt format, system instructions, and model version all affect results. If they are not captured at execution time, the evaluation cannot be reproduced or compared.
    • Cherry-picking risk. Without tamper-evident records of every run, there is no way to confirm that unfavorable results were not excluded.
    • No cryptographic binding. Published benchmarks rely on trust. Nothing ties the claimed score to the actual execution that produced it.

    What gets certified

    A Certified Execution Record for an evaluation run captures everything needed to reproduce and verify the result:

    • Test inputs including the exact prompt, system instructions, and test data used for that run
    • Model parameters including model version, temperature, maxTokens, and any configuration. This makes results comparable across runs, versions, and teams.
    • Model outputs including the complete response, finish reason, and token usage
    • Tool calls if the evaluation involves function calling or retrieval
    • Scoring metadata when the evaluation includes automated scoring, classification, or grading criteria

    All fields are sealed with a SHA-256 hash. The result becomes a verifiable artifact bound to the exact conditions that produced it, not a self-reported number in a spreadsheet.

    Evaluation workflows

    Model selection and comparison

    Compare models with verifiable evidence. Every test run produces a CER that proves exactly what was tested, with what parameters, and what the model returned.

    Safety and alignment testing

    Certify that safety evaluations actually happened. Produce tamper-evident records of red-team prompts, guardrail tests, and alignment checks.

    Regression testing

    Track model behavior across versions with verifiable test results. Detect drift with evidence, not assumptions.

    Benchmark publication

    Publish benchmark results backed by independently verifiable execution records. Let reviewers confirm the results without re-running the tests.

    Who benefits

    Procurement and vendor evaluation

    Evaluate AI vendors with independently verifiable test results. Require CERs as part of vendor selection and ongoing review.

    ML engineering teams

    Track model performance across versions with verifiable records. Make deployment decisions backed by sealed evidence, not internal notes.

    Compliance teams

    Demonstrate that required testing was performed with tamper-evident evidence. Satisfy audit obligations for AI system validation.

    AI research teams

    Publish evaluation results with cryptographic evidence. Increase credibility and reproducibility of published findings.

    How it works

    1. Wrap each test run. Add certifyDecision() to your evaluation pipeline. Each execution produces a sealed CER with full parameter and result context.
    2. Parameters are captured automatically. Model version, temperature, system instructions, and test data are all bound into the record at execution time.
    3. Request attestation. Submit CERs to an independent attestation node for Ed25519-signed receipts.
    4. Share verifiable results. Distribute CERs alongside evaluation reports. Anyone can verify the results using verify.nexart.io, without trusting the team that ran the tests.

    Cite this definition

    "Certified AI evaluation is the practice of producing tamper-evident, independently verifiable execution records for AI test runs, benchmarks, and model comparisons, replacing self-reported results with cryptographic evidence."

    , Certified AI Evaluation, NexArt Glossary. https://nexart.io/glossary#certified-ai-evaluation

    Make your AI evaluations verifiable

    Replace self-reported test results with Certified Execution Records. Every run becomes independently verifiable evidence.

    Quickstart guideVerify a recordView pricingContact us

    Related