Page 86 - CyberSecurityEssentialsEnglish

TABLE 2-12: LEVELS OF CONTROL OVER THE SHAPE OF THE OUTPUT

Level The benefit The limit

Describing the format in The formatting may change
Easy and flexible for people.
natural language between attempts.

Communicates the pattern It may copy details of the
An example of the output
more effectively. example or neglect other cases.

Easier to process It does not guarantee the
Explicit JSON
programmatically. structure or the values.

Reveals missing fields and
Schema validation It needs code and a retry policy.
wrong types.

Enforces the structure in It does not guarantee that the
Structured outputs
systems that support it. content is correct or true.

A Simplified Example of an Extraction Schema

Suppose the system extracts follow-up tasks from meeting minutes. The schema might state
that each item contains: task as text, owner as text or null, due_date as a date or null, status
with a value from a closed list, evidence as a quotation, and needs_review as a boolean. This

description prevents arbitrary fields from appearing and makes the handling of missing
information explicit.

After the output is received, the program checks the types and the permitted values. If
validation fails, the output can be rejected or retried with a specific error message. If the

structure passes, it is still necessary to check that the quotation exists and that the owner or
the date was not inferred without evidence. This is the boundary between correctness of

form and correctness of meaning.

For the non-programming reader, the practical idea is the same: do not stop at “put it in a
table.” Define the column names, the permitted values, how missing information is
represented, and when a case should be raised for review. The more an output becomes

part of an automated process, the stricter these rules must become.

⚠ Warning: a correct structure does not mean correct content
An output may match a JSON schema perfectly and still contain a wrong value or an
unsupported inference. Structural validation prevents one class of error. It does not
replace verification against the source and the logic.

85