Page 138 - CyberSecurityEssentialsEnglish

TABLE 3-13: THE RIGHT MEASURE FOR EACH TYPE OF TASK

Task Useful quantitative measures Necessary qualitative review

Are the errors concentrated in a
Accuracy, precision, recall, F1,
Classification sensitive class or a particular
and the confusion matrix.
group?

Does an error in a number, a
Accuracy per field, completeness,
Extraction name, or a unit change the
and the automatic pass rate.
decision?

Recall and precision at k, the rank
Retrieval Does the passage support the
of the result, and citation
systems claim, and is the version current?
faithfulness.

Did it preserve the meaning and
Writing or A rubric covering purpose,
the exceptions without adding
summarizing coverage, constraints, and time.
facts?

Agents and The proportion of completed Is the action authorized,
execution steps, tool errors, and rollbacks. stoppable, and explainable?

Using a Model as a Judge: Useful, but Not a Final Verdict
A language model can be used to evaluate many outputs against a rubric, but it may be
affected by the order of the answers, by length, and by a preference for a style or a model
resembling itself. Zheng and colleagues documented these limits while also showing the

value of an automated judge for working at scale [10]. When measuring truthfulness or
resistance to misleading answers, benchmarks designed for that purpose can help, while

local testing and the context of the task remain decisive [21].

When using a model as a judge, shuffle the order of the answers, hide the identity of the
product, define the criterion, repeat the evaluation, ask for brief evidence, and compare
against a human sample. Do not use its judgment as the sole criterion in a high-impact task,

and do not let the generating model, the judging model, and the reference source be the
same without controls.

⚠ Warning: controls when a model acts as a judge
Shuffle the order of the answers, hide the name of the model, use a defined rubric,
repeat the evaluation, compare against a human sample, and check for bias toward

length and style.

137