Reports
Is the system getting better, is it earning, and where is everything right now?
Sounds like the writer agrees with people least (score 0.40, target 0.60). Until it improves, its results keep going to a person.
Agreement with people, check by check64 labelled articles
| Check | Agreement score | Same answer | Blocked good work | Let bad work through |
|---|---|---|---|---|
Sounds like the writer How close the style is to the writer's real articles | 0.40 | 71.9% | 17.2% | 10.9% |
No AI tells Overused AI words and formulaic phrasing | 0.66 | 84.4% | 9.4% | 6.3% |
Original enough Not too close to our own pages or the top search results | 0.80 | 90.6% | 7.8% | 1.6% |
Facts have sources Every fact points to a source | 0.86 | 93.8% | 1.6% | 4.7% |
SEO and links Keyword, recipe markup, links, image descriptions | 0.87 | 93.8% | 6.3% | 0% |
Cook's notes attached Real test notes and the cook's own photo | 1.00 | 100% | 0% | 0% |
How to read this
- Agreement score (Cohen's kappa) corrects for agreeing by luck. 0.60 or higher is the target; the line on each bar marks it.
- Blocked good work costs reviewer time. Let bad work through risks publishing a weak page. They aren't equally bad, so they're never added together.
- Scores need at least 20 labelled articles per check; below that the reason is shown instead of a number.