Five engines could read a note into fields: keyword rules, a trained classifier, a trained tagger, a 1 GB model running in the browser, and a cloud model. Thirty realistic notes with the right reading for each went through every engine with one scorer. Thirteen models in all, across two providers, for about one dollar. The table shows the rules, the browser model and three of the cloud models.
| Engine | Kind right, of 30 | Written time right, of 30 | Invented people | Seconds a note | Cost per 1,000 |
| Keyword rules | 25 | 26 | 19 | 0.01 | free |
| 1 GB model in the browser | 18 | 21 | 0 | 15 | free |
| Cheapest cloud model | 25 | 24 | 2 | 1.4 | $0.06 |
| Mid cloud model | 30 | 29 | 1 | 1.6 | $0.80 |
| Largest cloud model | 30 | 30 | 1 | 2.4 | $3.20 |
The rules treated every capitalized word as a name, and the 1 GB browser model copied the prompt's example sentence for seventeen notes. When a flagship model lost to cheaper ones I challenged the result and checked the harness: the prompt was ambiguous, and it compared models at different reasoning levels without recording it. The table shows the rerun, with the prompt rewritten and each model's reasoning level recorded. The mid-priced cloud model was chosen: as accurate as the largest on the kind of note, for a quarter of the cost.