From intended use to a documented decision.
A campaign binds one system version to a frozen evaluation definition, observed behavior and human review.
Evaluate the declared system.
The system boundary includes prompts, model configuration, retrieval, tools, application logic, policy rules, output filters, human escalation and declared external services. A foundation-model score alone does not describe that whole workflow.
The local prototype runs a synthetic servicing application. Connecting a customer application and agreeing representative tests remain separate evaluation work.
- 1
Define the intended use
Record the owner, operating context and system boundary.
- 2
Freeze the system version
Bind prompts, model configuration, retrieval, tools and escalation.
- 3
Set the evaluation campaign
Choose cases, checks, thresholds and the policy revision.
- 4
Capture observed behavior
Keep results, outputs, failures and evaluator provenance.
- 5
Compare the change
Separate new failures, resolved failures and changed coverage.
- 6
Record human review
Document the decision, rationale, conditions and unresolved findings.
- 7
Export the evidence
Retain a readable report and machine-readable artifacts.
Preserve uncertainty.
- PASS
- Every required check completed and satisfied the frozen policy.
- FAIL
- An observed blocking failure remains.
- REVIEW_REQUIRED
- The evidence requires human interpretation, including model-judge disagreement.
- INSUFFICIENT_EVIDENCE
- Required evidence is missing, invalid, interrupted or not evaluated.
When failures and uncertainty coexist, the report retains both. A later human decision cannot rewrite the original observations or gate.
Evidence supports a decision. It does not make one.
Results apply to the specified system version, evaluation policy and tested cases. Missing evidence remains visible. Human review is a separate record.
Nikxius does not determine regulatory compliance, approve financial AI or establish that a system is safe. Representative domain coverage and institutional reviewer acceptance remain unvalidated.
Read the trust boundariesInspect the workflow yourself.
The synthetic demo includes corrected failures and an intentional hardship-routing regression.
Run the demo