Scoreboard
A wrong value is worse than a flag. Wrong is the number we drive to zero. Flagged is the human workload. Measured against golden.json, Claude Sonnet 5.5, cached results.
All fields
64 fields
- Correct
- 60
- Flagged
- 4
- Wrong
- 0
Conflict cases
10 fields
- Correct
- 6
- Flagged
- 4
- Wrong
- 0
Clean cases
54 fields
- Correct
- 54
- Flagged
- 0
- Wrong
- 0
Per document
| Document | Correct | Flagged | Wrong | Cost | Time |
|---|---|---|---|---|---|
| doc-01 home | 7 | 1 | 0 | $0.0141 | 7.0 s |
| doc-02 home | 8 | 0 | 0 | $0.0130 | 6.2 s |
| doc-03 motor | 7 | 1 | 0 | $0.0141 | 6.3 s |
| doc-04 motor | 8 | 0 | 0 | $0.0152 | 6.7 s |
| doc-05 home | 7 | 1 | 0 | $0.0146 | 6.4 s |
| doc-06 motor | 7 | 1 | 0 | $0.0133 | 6.3 s |
| doc-07 home | 8 | 0 | 0 | $0.0136 | 6.0 s |
| doc-08 motor | 8 | 0 | 0 | $0.0127 | 5.9 s |
| Total | 60 | 4 | 0 | $0.1106 | 50.9 s |
Cost is API-equivalent: tokens × the model's price per token. Time is the extraction run per document.
Rules in use
No rules yet. Every disagreement is flagged until a reviewer adds a rule.