A specialist wants the appointment-booking topic to refuse any request to book a rabies-booster slot more than 30 days out, every time, with no exceptions. Rather than adding an instruction telling the model to check the date, what should the specialist configure?
Select an answer to reveal the explanation.
Short Explanation
An instruction is a suggestion you're hoping the model takes seriously; a filter tied to a variable is a rule the platform enforces before the model even gets a say. If the answer must be right every single time, you don't ask nicely, you make it structurally impossible to get wrong.
Full Explanation
Filters and variables exist precisely for rules that must hold every time, independent of how the model reasons about a given utterance. Binding a filter on the booking action to a variable carrying the 30-day threshold means the platform evaluates the condition deterministically before the action can execute, so the outcome does not depend on model interpretation. Restating the rule more emphatically in an instruction is still probabilistic guidance; a model can and eventually will misapply it under an unusual phrasing, which is exactly the failure mode determinism is meant to close. Splitting the rule into a second topic changes classification, not enforcement, and a topic boundary is a poor substitute for a hard business rule since classification itself is probabilistic. A guardrail phrased as an apology instruction is still an instruction-level control, not a structural one, and does nothing to stop the action from actually running. For the exam, remember that filters and variables are the tool for anything that must never vary, while instructions and guardrails shape behaviour that is allowed to vary with context.