Evaluation harness

An evaluation harness is a test environment for repeatedly measuring an AI system's quality. It assesses the system using a fixed collection of tasks with known correct answers.

How it works

First, a test set is created from real, anonymised cases with verified outcomes. The AI system processes each case, and the harness compares its response to the expected result. This generates key metrics like hit rate, false assessment rates, and time per case. The harness runs again with every change, such as a new model, prompt or data source.

For example, a SOC wants to switch to a new model version. The harness has both versions evaluate 500 previous Cases. The new version is faster but classifies more real attacks as harmless. The team stays with the old version for now and adjusts the instructions.

What to look out for

• The test set must match your actual tasks. General benchmarks reveal little about performance in a SOC.
• Weight errors according to their consequences. A missed attack is more serious than a false alarm.
• Continuously expand the test set with new and challenging cases.
• Make sure that test cases do not contain any confidential data.

Why it matters

AI systems often change their behaviour unexpectedly when the model or instructions are altered. Without measurement, a drop in performance is only noticed during live operation. A harness makes these changes visible before deployment.

Typical mistakes

Frequently, the test set consists only of simple cases. The system then appears more effective than it is in daily operations. A second mistake is a test set that is never updated, even though attacks evolve.

How we implement it

Our Agentic AI operates in the SOC within the scope of your mandate. High-impact interventions remain subject to approval by our analysts.