Q30 of 60 — CCA Practice Exam

Structured Data Extraction

Your structured data extraction pipeline processes 500 legal contracts per night, pulling out clause summaries, liability caps, and termination conditions. Post-deployment review reveals that in roughly 12% of contracts, Claude generates liability cap figures or termination dates that do not appear anywhere in the source document — a hallucination rate that is causing downstream compliance failures. The contracts average 30,000 tokens each. What is the most effective architectural change to reduce fabricated field values in the extracted output?

A. Require Claude to extract a verbatim quote from the contract supporting each target field before populating the output, and emit a null sentinel for any field where no supporting quote can be found.

B. Run each contract through the extraction prompt three times and use majority-vote logic to select the most frequently returned field values, discarding any outlier responses that differ across runs.

C. Add a system prompt instruction telling Claude to be conservative and only include values it is highly confident about, flagging uncertain fields with a numeric confidence score between 0 and 1.

D. Increase the output token budget for each extraction call so Claude has more room to reason carefully over the full contract text before committing to specific field values.