Retrieval and evidence
Measure relevant evidence recall, citation correctness, permission filtering, and response to missing sources.
Evaluation methodology
How do you know an enterprise AI system is ready?
A model benchmark is not an acceptance test. Production readiness must be evaluated across evidence retrieval, output quality, tool behaviour, workflow reliability, human control, and operating outcomes.
By Onyx Devs Lab · Published · Updated
CapabilityCan it perform the task?
ControlDoes it fail safely?
OutcomeDoes the workflow improve?
Six evaluation layers
Each layer answers a different production question.
Output and decisions
Use domain rubrics for factuality, completeness, policy compliance, calibration, and appropriate refusal.
Tools and operations
Test argument construction, authorisation, idempotency, retries, post-action verification, latency, and cost.
From test set to operating evidence
Offline scores are necessary but insufficient.
Representative cases
Build sets from common work, high-impact edge cases, historical failures, and adversarial inputs.
Release gates
Define thresholds by risk tier and prevent averages from hiding severe failures.
Production feedback
Capture corrections, escalation, abandonment, overrides, and downstream outcomes without treating every user action as ground truth.
Delivery method
Define the consequential task
State what the system may influence and who bears the consequence.
Build an evidence-backed test set
Record inputs, expected evidence, rubric, risk tier, and acceptable alternatives.
Evaluate the full workflow
Include retrieval, model, tools, permissions, interface, and human handoff.
Monitor after release
Track drift, failures, cost, latency, overrides, and the agreed business metric.
Frequently asked questions
Is model accuracy enough?
No. A model can score well while retrieval misses protected evidence, tools perform the wrong action, or the workflow fails to provide human control.
How large should an evaluation set be?
It should cover the decision and risk space rather than chase an arbitrary count. Begin with high-frequency and high-consequence cases, then expand from observed failures.
Can user feedback replace expert evaluation?
No. User feedback is useful operational evidence, but many users cannot verify hidden factual, policy, permission, or downstream errors.
Start with one concrete operating problem.
We will first clarify the workflow, evidence, data boundary, and outcome standard before recommending the next step.
Contact Onyx