Human evaluation around defined AI policies and test sets.

Nardux can assess bounded workflows for reviewing approved prompts, outputs, reports and evaluator labels. Model safety, security testing and deployment decisions remain outside the implied scope.

Evaluation loses meaning when context or reference changes silently.

01

Rubrics stay abstract

Evaluators cannot apply high-level prohibited-use or output principles consistently.

02

Versions are mixed

Results combine different models, prompts, system settings or policy versions.

03

Disagreement is flattened

Automation and humans differ, but the reason and reference decision are not analyzed.

04

Test sets imply too much

A bounded sample is treated as proof of universal safety, fairness or compliance.

Make the evaluation reproducible and appropriately narrow.

RUBRIC & TEST SET

Define the reference

  • Policy dimensions and severity
  • Model, version and configuration
  • Curated case sources and boundaries
  • Expected labels and reason codes
  • Edge-case and disagreement process
HUMAN REVIEW & QUALITY

Preserve accountability

  • Reviewer competence by modality
  • Independent review and calibration
  • False-positive and false-negative analysis
  • Failure taxonomy and escalation
  • Versioned findings and limitations
NOT IMPLIED

Keep pilot claims narrow

  • No comprehensive model-safety certification
  • No cybersecurity or adversarial red teaming
  • No guaranteed regulatory compliance
  • No autonomous model or deployment changes
  • No unapproved high-impact or specialist evaluation

Every result belongs to a defined test context.

Reporting should identify the rubric, test-set construction, model and version, modalities, evaluator process, reference outcome and disagreement method. Results should distinguish false positives, false negatives, policy ambiguity and legitimate judgment ranges.

A pilot report supports product and human-review decisions within the tested scope. It is not proof that the model is safe across users, prompts, languages, versions or high-impact uses.

Approve the rubric and data before evaluating outputs.

  1. 01 / ASSESS

    Define the use case and risks

    Review the product, models, users, modalities, policies, test purpose, data and specialist gates.

    Explore consulting
  2. 02 / BUILD

    Operationalize the rubric

    Create labels, severity, examples, reasons, disagreement and escalation guidance.

    Explore policy support
  3. 03 / PILOT

    Evaluate a controlled set

    Run human review and quality analysis only after tool, data and reviewer readiness.

    Explore quality assurance

Three questions to bring to discovery.

  1. 01

    What exact decision should the evaluation support?

    Define the product, policy, model version, modalities, population and intended use of results.

  2. 02

    How was the test set created and approved?

    Identify sources, rights, personal data, sensitive categories, representativeness and exclusions.

  3. 03

    Who resolves evaluator disagreement?

    Name the policy reference, final owner, escalation route and change process before scoring begins.

Questions before an evaluation pilot.

Does Nardux certify AI models as safe?

No. A bounded human evaluation can describe findings for an approved rubric, test set, model version and period. It cannot establish universal safety or regulatory compliance.

Do you perform AI red teaming?

Not as an implied service. Cybersecurity or adversarial red teaming requires separate specialist expertise, tooling, authorization and risk controls.

Can client or user data be sent to third-party AI tools?

Not without explicit approval of the use case, vendor, data flow, terms, retention, access, security, privacy and human-oversight model. Sensitive data is not assumed suitable.

Can findings be compared across model versions?

Only when the test method, rubric, cases, settings and limitations support a meaningful comparison. Version and configuration changes must remain visible in reporting.

Start with the decision, rubric and bounded test set.

Share business-level context only. Do not submit prompts, outputs, datasets or confidential model information through the public form.

Discuss an Evaluation Pilot