Test the System

Acme now gets the right answer. One successful record proves the concept, not the system. Rules interact, retrieval changes, and a fix for one workflow can break another.

Turn every known failure into a question

A targeted golden question fails without one rule and passes with it. The Acme authority conflict becomes a permanent test. A broad question exercises several rules together and catches interactions the targeted test misses.

Grade reasoning before prose

Check field selection, rule compliance, exclusions, date disclosure, permission awareness, refusal quality, and stability across runs. A fluent answer that reaches the right number by accident still fails.

Record what produced the result

Every test run needs the guidance version and loaded rule IDs. Without them, a changed answer cannot be traced to a revision. Deploy rule text and retrieval logic separately so the diff has one plausible cause.

Let production mistakes grow the suite

Start with one question per rule and a few broad checks. When a user corrects an answer, reproduce it, write the missing rule, and add the question. The suite grows from real failures instead of imagined coverage.

Build a small defensible regression run in the GTM Lab.

Reference

• Golden-question testing and rubric →
• Thirty-question starter set →