When only 5% of replies fail, a judge that says "pass" to everything is 95% accurate and useless. Accuracy rewards agreeing with the majority, and in a good product the majority is fine. For rare failures, look at:
Recall (the TPR from the last lesson): the share of real failures the judge catches.
Precision: of the replies the judge flags, the share that really fail. Low precision means reviewers waste time on false alarms.
Two judges graded the same 400 replies. One looks better on accuracy. Find out which one is actually useful, and when.
Your task
Compute accuracy, precision and recall for both judges.
For three situations, choose the judge you would deploy.
Sign in to practise
The lesson is free to read. A free account lets you talk to the practice bots, run your code and save your score. You will come back to this lesson.
How this is checked: When you submit, your script runs and the server compares what it prints with the right values.
2
Label each one
Which judge would you deploy?
0 of 3 labelled
1Failures are wrong refund promises; every miss reaches a customer.
2Failures are cosmetic, and every flag costs an hour of a senior reviewer's time.
3A new product where about 40% of replies fail. You want the most correct verdicts overall.
How this is checked: Each row has a right label. All 3 must match. This is a harder check: it earns the second 50 points once the core checks pass.
Core checks earn 50 points and the harder checks the other 50. You can submit as many times as you like.
About this lesson
Precision and recall for an LLM judge with rare failures: the accuracy trap
AI Testing for QA Engineers·Intermediate·Judges and guardrails as classifiers
How it works
When only a small share of replies fail, a judge that passes everything has high accuracy and is useless. Accuracy rewards agreeing with the majority, and in a good product the majority is fine.
For rare failures, two other numbers matter. Recall is the share of real failures the judge catches. Precision is the share of the judge's flags that are real failures; low precision wastes reviewers' time on false alarms. Which matters more depends on what happens after a flag.
About the assignment
Compute accuracy, precision and recall for two judges graded on the same 400 replies, and choose a judge for three situations.
What you will do
1Build each judge's confusion matrix from the counts.
2Compute the three metrics for both.
3For each situation, ask what a miss costs and what a false alarm costs.
Common mistakes
Choosing by accuracy
At a low failure rate, accuracy barely moves whether the judge works or not.
Ignoring reviewer time
A judge that flags too much can drown a review team.
In a real test suite
Teams choose the judge threshold by the cost of each mistake. A safety screen favours recall; a queue sent to busy reviewers favours precision.
It also helps to look at the replies each judge gets wrong. A judge that misses one specific kind of failure every time can sometimes be fixed with a single clause in its rubric, which is cheaper than replacing it.
