Free Claude Certified Architect Foundations (CCAR-F) Sample Questions

D. The model picks the right action and supplies every field. It is only misreading the unit of amount, because the schema says nothing beyond "integer". The only fix is a parameter description that states integer cents and gives a concrete example, since the model relies on tool and parameter descriptions to form its inputs. Here the calls succeed with a valid but wrong value, so the ambiguity must be removed at the interface itself. Why the others are wrong: An enum on action helps when the agent sends invalid or misspelled action strings, but every action value here was already valid and correct. Structured errors help the agent recover when calls fail, but these calls succeed with a well-typed wrong value, so no error is ever returned. Splitting the tool is the right fix when the agent picks the wrong action or mixes up fields between modes, but here action selection is always correct, and a separate adjust_billing tool would still carry an undocumented amount unit. Source: CCAR-F exam guide, Task Statement 2.1.

C. Aggregate accuracy can hide weak performance in a specific document segment, so a labeled set should be broken down by document type as well as by field. Here that breakdown shows the model's confidence is miscalibrated for Provider X's layout: its wrong answers still score above 0.97. A confidence cutoff cannot separate good from bad values in that segment, so the whole segment needs human review. The well-calibrated segments can keep auto-approval. Why the others are wrong: A higher global threshold would not catch Provider X errors, because they score above 0.97, and it would add needless review to accurate providers; this change fits when errors cluster just above the current cutoff. Routing low-confidence fields to review is sound when confidence tracks correctness, but the Provider X errors are high-confidence and would pass straight through this rule. Stratified sampling is the right way to keep measuring error rates in production, but here it would knowingly auto-approve a segment already shown to be only 81% accurate. Source: CCAR-F exam guide, Task Statement 5.5.

D. Only a subagent's final message returns to the coordinator, so everything placed in that message lands in the coordinator's context. With 2,300 locations, returning the full list inline floods the main agent. Persisting the list to a file and returning a lightweight reference keeps the main context lean. Every location stays retrievable for per-package edit planning, which is the requirement the stem imposes. Why the others are wrong: Parallel subagents would still return the same 2,300 locations to the coordinator in total, so the context flood remains; fan-out helps when a single trace is too slow or too broad for one subagent. A condensed summary would keep the main context lean but would drop most site locations the later edit planning needs; it would suit a case where the coordinator only needs an overview. The subagent already stayed in scope and found everything, so missing context was not the problem; passing prior conversation helps when a subagent ignores constraints it was never given. Source: CCAR-F exam guide, Task Statement 1.2 / 1.3.