Artificial intelligence · March 6, 2026 · Gynvix · 1 min read
Keep a small evaluation set
Ten real questions, with the expected source, will tell you more than a leaderboard number.
Questions from the work
Collect questions people already ask. For each one, record the document that should have answered it.
What to score
Did the answer cite that document? Did it refuse when the document was absent? Did it invent a step that is not in the text?
When to run it
Run the set when the documents change or the prompt changes. A green build that never sees those questions is not a check.