Claude Certified Architect – Foundations Practice Question 31: Evaluator Optimizer Loop
Claude Certified Architect, Foundations · Test 4 · question 1 of 10
A localization team uses Claude to translate marketing copy. First-pass translations are fluent but often miss brand-tone guidelines and idiom nuances, and reviewers must send them back several times. The team has a written rubric (tone, terminology, idiom handling) and wants to improve quality automatically before a human sees the text. Which design best fits?
Show answer and explanation
Answer:
C is the evaluator-optimizer workflow: one LLM call produces a response and another provides evaluation and feedback in a loop. It works well when there are clear evaluation criteria (here, the rubric) and when iterative refinement gives measurable value. A retry cap keeps cost and latency bounded.
A is parallelization by voting or sampling, but picking the longest output has no relation to quality or to the rubric. B is a decomposition that loses cross-sentence context and tone consistency and has no feedback step. D is not a Claude API option for this scenario (fine-tuning is not generally offered for current Claude models) and would not give per-document rubric checks anyway.
Reference: Building effective agents