Writes grading notes with acceptable alternatives and named failure modes
The checklist rows reached, the items, the acceptable alternatives and the named failure modes
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 | 32% | Not shown |
| gpt-6-luna | Low cost | 98% | Mastered |
| claude-haiku-5-5 | Low cost | 95% | 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 checklist rows reached, the items, the acceptable alternatives and the named failure modes
How it is assessed: A grader can credit a correct answer that differs from the reference and name the expected failures
Related knobits
Builds on
All 15 knobits of Write and review test items.
Source: Graded on its own by sa-item-authoring