Recognises disagreeing severity, threshold and failure action
severity, threshold and on_failure telling different stories: a warning that rejects rows, an error that only reports, a threshold below 100 % on a check that rejects every failing row anyway. Somebody disagreed and nobody decided.
Can AI do this?
People still lead Few or none of the tested AI models showed they master this. 0 of 4 models in the current benchmark master it.
| Model | Tier | Probability of mastery | Status |
|---|---|---|---|
| gemini-4-argon | Flagship | 83% | Not shown |
| gemini-3.1-pro-preview* | Flagship | 99% | Mastered |
| gpt-6-astra* | Flagship | 99% | Mastered |
| claude-fable-5-1* | Flagship | 100% | Mastered |
| gemini-3.8-flash* | Mid tier | 100% | Mastered |
| gpt-6.1-sol* | Mid tier | 99% | Mastered |
| claude-sonnet-5-5* | Mid tier | 99% | Mastered |
| gemini-3.5-flash-lite | Low cost | 33% | Not shown |
| gpt-6-luna | Low cost | 84% | Not shown |
| claude-haiku-5-5 | Low cost | 81% | Not shown |
* From the earlier benchmark of 2026-10: 10 data-modelling test tasks, 3 runs per task, graded by claude-opus-5-5. These models have not taken the current 123 tasks yet; the two methods differ, so compare their figures with care. They are left out of the “most models” counts and the best-model picks.
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 three settings made consistent
How it is assessed: Flags the deficient case and passes the acceptable one. Deficient: severity: warning, on_failure: reject Acceptable: severity: error, on_failure: quarantine, threshold 100 %; or severity: warning, on_failure: flag, threshold at the agreed rate (DQR-015, DQR-026)
Also known as: Severity, threshold and action that disagree (data quality rule)
Related knobits
Builds on
All 30 knobits of Derive data quality rules.
Source: Review checklist 03-dq-rules (data-quality), row 21; Kimball & Caserta 2004; Loshin 2011; formerly data-modelling K-721