AI & LLM Evals / Scenario

The Eval Gate That Never Fails

A DeepEval CI gate has been green for six weeks, including the release that shipped an obvious hallucination regression. Read the test code and find why the gate is theater.

Difficulty
Hard
Format
Scenario
Points
200
Estimate
14 min

// MISSION BRIEF

Your Mission

The team wired DeepEval into CI months ago and the badge has been green ever since. Last week a release started inventing refund policies, support tickets exploded, and the gate never blinked.

You have the test file and a passing CI log. Work out why this gate physically cannot fail, and what a real one looks like. Decisions are weighted.

// FIRST CONTACT

Battle teaser

First artifact

test_rag_quality.py

Why is faithfulness pinned at 1.00, even while the bot hallucinates policies?

  1. AThe threshold of 0.2 forces the score to round up
  2. BThe judge model is too small to detect hallucinations
  3. CFaithfulness always returns 1.0 unless the answer is empty
  4. DCircular grounding: retrieval_context is filled with the model's own answer, so every claim is 'supported' by itself; the metric must receive the chunks the retriever actually returned
Answers, scoring, hints, and the full battle stay sealed.

// SKILL TAGS

deepevalci-cdfaithfulnesseval-engineering