LLM evaluation: measure the capability actually required
Evaluate an LLM with behavior contracts, protected sets, OOD robustness, schema accuracy and baseline comparison.
Useful evaluation starts with the required capability and forbidden failures, not a generic benchmark.
A high average score can coexist with rare but unacceptable failures on the real mission.
The verdict must be bounded to the model, prompt, runtime, data and protocol that were actually measured.
What the architecture must make explicit
Evaluating an AI system binds a capability, an exact object, a protocol, observations and a decision rule. An isolated score is not a qualification.
Move from a keyword to an inspectable engineering decision
Contract
Define required behavior, forbidden failures and thresholds before testing.
Separate
Keep search, confirmation, protected data and Fresh Final apart to prevent contamination.
Measure
Record outputs, metrics, failures, latency and measured-subject identity.
Adjudicate
Interpret observations without letting the candidate producer impose the verdict.
A claim is only as strong as its binding
The gate must evaluate the exact output produced by the exact realization. Integrity, scientific validity and authority are related, but never interchangeable.
Questions that change the design
What does llm evaluation: measure the capability actually required require?
Useful evaluation starts with the required capability and forbidden failures, not a generic benchmark. Evaluating an AI system binds a capability, an exact object, a protocol, observations and a decision rule. An isolated score is not a qualification.
Which evidence should be retained?
At minimum: exact realization identity, inputs and outputs, protocol version, measurements, failure cases and the decision rule. The verdict must be bounded to the model, prompt, runtime, data and protocol that were actually measured.
Does AMF prescribe a particular model?
No. AMF treats the model, runtime, data, tools and deterministic components as candidate materials. Their retention depends on measured contribution and declared constraints.