CORD-v2 test receipt · read by Qwen3.5-4B (QLoRA) with best-of-4 and the accounting checker
Receipt photo

How the checker chose
The model wrote one normal reading and four sampled ones. The checker keeps the reading that passes the most accounting rules; the model's confidence only breaks ties.
- Greedy83% rulespicked
- Sample 183% rules
- Sample 283% rules
- Sample 383% rules
- Sample 483% rules
Extracted fields
| Item | Qty | Unit price | Amount | Remove |
|---|---|---|---|---|
Accounting checks
goes to human review
- Line 1: qty × unit price = amountexpected 432,000 · found 432,000
- Line 4: qty × unit price = amountexpected 222,000 · found 222,000
- Line 7: qty × unit price = amountexpected 60,000 · found 60,000
- Line 8: qty × unit price = amountexpected 70,000 · found 70,000
- Items add up to the subtotalexpected 1,451,000 · found 1,343,000
- Subtotal − discount + tax + service = totalexpected 1,565,938 · found 1,565,938
The rules recheck as you edit, using a TypeScript port of the Python checker that agrees with it on all 277 rule results of the test set. Amounts within 1% are treated as equal, to allow for rounding.