QABattle

AI & LLM Evals / Scenario

Pick the Detector

Four model outputs, four different flavors of wrong. Match each to the detection method that would actually catch it, and spot the confident fabrication.

Difficulty
Hard
Format
Scenario
Points
200
Estimate
15 min

// MISSION BRIEF

Your Mission

A medical-information assistant produced four answers under review. Each is wrong in a different way, or right. Your job: choose the detection method that fits each case and flag the most dangerous output.

There is no one-size detector. Match the method to the failure.

// FIRST CONTACT

Battle teaser

First artifact

Four outputs under review

Output A (context says 400, answer says 800) is best caught by:

  1. ASelf-consistency alone
  2. BAnswer length
  3. CFaithfulness/groundedness (or NLI entailment), the claim contradicts the retrieved context
  4. DCitation formatting checks
Answers, scoring, hints, and the full battle stay sealed.

// SKILL TAGS

ai-evalshallucinationfaithfulnessself-consistencycitations