Page 151 - CyberSecurityEssentialsEnglish

Common
Term Plain-language explanation
abbreviation

Benchmark Test material leaking into training data, so that a score
-
contamination overstates real generalization.

A document summarizing intended uses, testing, limits, and
Model card -
measures.

A document some providers publish about the whole
System card -
system and its testing; there is no single global format.

A document describing why a dataset was collected, how it
Datasheet -
is composed, and how it should be used.

Reference Representative local cases used to compare tools and
-
evaluation set detect change.

A small repeated test rerun after an update to detect a
Regression test -
change in behavior.

A classification of failure types that guides the remedy
Error taxonomy -
rather than only the score.

How well a stated confidence or probability matches actual
Calibration -
accuracy.

How well performance holds up when the input changes:
Robustness -
different phrasing, noise, a longer file, or an unusual case.

The ability of a system to decline to claim when the
Abstention -
information is insufficient.

Choosing a different model or tool according to the type of
Model routing -
task.

Putting the same task to systems from different providers
Triangulation - and treating the point of disagreement, not the agreement,
as the signal that needs checking.

Total cost of The cost of purchase, integration, operation, training,
TCO
ownership support, and exit.

The ability to repeat an experiment under documented
Reproducibility -
conditions and obtain a similar result.

“The right tool is not the best known or the largest. It is the system that succeeds at your

task, with your data, in your language, under your constraints, and whose results you can
explain, verify, and stop when you need to.”

150