Recognises a reference answer failing its own rubric
The reference, graded by the rubric, would not get full marks.
Can AI do this?
AI does this well Most of the tested AI models master this. 3 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 | 30% | Not shown |
| gpt-6-luna | Low cost | 98% | Mastered |
| claude-haiku-5-5 | Low cost | 98% | Mastered |
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 criterion it fails and the fix
How it is assessed: Flags the deficient case and passes the acceptable one. Deficient: Criterion I2 requires CU-07; the reference answer never mentions it Acceptable: The reference graded against every criterion before release
Also known as: Reference answer failing its own rubric
Related knobits
Builds on
All 15 knobits of Write and review test items.
Source: Review checklist 02-test-items, row 3 (Standards 4)