Page 153 - CyberSecurityEssentialsEnglish

Steps: Run the same prompt five times without changing anything. Record what stays
fixed and what changes. Then classify each change as content, form, or fact.
Output: A comparison table of the five runs.
Assessment criterion: Useful variation is separated from error, and variation on its own is
not treated as evidence of poor quality.

↻ Exercise 3: Diagnosing the intervention

Goal: Avoid applying one remedy to every kind of failure.
Required: Five errors you have already encountered, with their outputs. Allow about 30
minutes.
Steps: Describe each error precisely. Then decide whether it is addressed by the prompt,
the context, retrieval, a tool, structured outputs, fine-tuning, or a workflow. Write the

reason for each choice.
Output: A decision map linking each error to one intervention.
Assessment criterion: Every intervention is tied to a stated cause rather than to habit.

↻ Exercise 4: Testing Arabic
Goal: Measure performance in Arabic inside a real context rather than from a
benchmark.
Required: One Arabic task and two tools you can compare. Allow about 60 minutes.

Steps: Build twelve cases spread across standard Arabic, dialect, terminology, numbers,
and layout. Run them on both tools without knowing which is which. Then classify the
errors instead of scoring overall.
Output: An error report with the classification.
Assessment criterion: Language correctness and factual correctness are not merged into
a single score.

↻ Exercise 5: An adoption card
Goal: Reach a decision that is written down and can be defended later.
Required: One tool under consideration and public or anonymized data for the trial.
Allow about 45 minutes.
Steps: Apply the mandatory gates first and reject the tool if any gate fails. Then score the
weighted factors. Then define the scope of use, the prohibited cases, the review owner,
the fallback, and the re-evaluation date.

Output: A one-page decision card.
Assessment criterion: The gates were applied before the score, and the card names a re-
evaluation date rather than treating the choice as permanent.
Warning: Use public or anonymized data during the trial. Do not send real personal or
confidential material to a tool you have not yet approved.
References and Selected Resources

I. Scientific and Technical References
[1] Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y., June 22, 2023), Opinion and Order
on Sanctions

152