Recognises an answer given away in the prompt
The prompt names the defect the item is meant to observe.
Can AI do this?
AI does this partly Some of the tested AI models master this; others do not show it. 2 of 4 models in the current benchmark master it.
| Model | Tier | Probability of mastery | Status |
|---|---|---|---|
| gemini-4-argon | Flagship | 98% | Mastered |
| gemini-3.5-flash-lite | Low cost | 24% | Not shown |
| gpt-6-luna | Low cost | 98% | Mastered |
| claude-haiku-5-5 | Low cost | 52% | Not shown |
Mastered means a probability of at least 90% that the model has this knobit, estimated from its graded answers. “Not shown” means the answers do not settle it either way. How the models are tested.
What it looks like in practice
Expected result: The cue removed
How it is assessed: Flags the deficient case and passes the acceptable one. Deficient: "Note: CU-03 and CU-08 are the same person." Acceptable: The prompt states the rules; the data shows the case
Also known as: Answer given away in the prompt
Related knobits
Builds on
All 15 knobits of Write and review test items.
Source: Review checklist 02-test-items, row 4 (Messick; Standards 4)