Page 111 - CyberSecurityEssentialsEnglish
P. 111
computation, summarization, and the creation of a visual chart. Do not expect a product to
be strong in all of these merely because it accepts the same file.
✓ Practical tip: draw the task as a chain
Draw the task as a chain of five to seven steps. Write one word above each step:
generation, retrieval, extraction, computation, or execution. If more than one category
appears, your real choice is the design of a system or a workflow, not the selection of a
single model.
The Impact of Failure and the Criterion of Success
Errors are not equal. A poor suggestion for a headline can be spotted at once, while an error
in a treatment recommendation, a contractual clause, or the advanced screening of a job
application may cause serious harm. This chapter uses the phrase impact of failure to
describe the severity of the outcome if the system gets it wrong or if someone uses it out of
place. The higher the impact, the stronger the source you need, the fewer the permissions,
the wider the testing, and the more specialized the human review.
Before comparing products, write four sentences: I will consider the output successful
when...; the most dangerous possible failure is...; I will verify by...; and if the system fails I
will use... . If you cannot complete these sentences, the task itself has not yet been defined.
▣ Practical example: extracting invoice fields
Context: a team wants to extract supplier names and invoice values from PDF files. The
task is not “summarize the invoices” but the extraction of fixed fields. The most suitable
approach: a document tool that supports OCR, tables, and a structured output.
Verification: review a stratified sample that includes both clean and poor files, measure
the accuracy of each field, and do not pass a row to the finance system if validation fails.
● Remember: Section summary
• Key idea: Begin with the verb, the output, and the evidence of success.
• Practical decision: Classify the task and define the impact of failure before comparing
names.
• Mistake to avoid: Choosing the best-known product and then forcing the task to adapt
to it.
3.2 What Does the Reader Actually Use?
◆ Key idea: the user does not deal with a model alone.
The user deals with a whole system that includes an interface, data, tools, policies, and a
workflow. Two experiences that use the same name may therefore differ a great deal.
110

