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.
- Greedy63% rulespicked
- Sample 150% rules
- Sample 263% rules
- Sample 363% rules
- Sample 463% rules
Extracted fields
| Item | Qty | Unit price | Amount | Remove |
|---|---|---|---|---|
Accounting checks
goes to human review
- Line 1: qty × unit price = amountexpected 3,500 · found 6,500
- Line 2: qty × unit price = amountexpected 3,000 · found 8,000
- Line 3: qty × unit price = amountexpected 18,000 · found 18,000
- Line 4: qty × unit price = amountexpected 0 · found 0
- Line 5: qty × unit price = amountexpected 0 · found 0
- Line 6: qty × unit price = amountexpected 5,000 · found 5,000
- Items add up to the subtotalexpected 37,500 · found 24,500
- Card charge = totalexpected 24,500 · found 24,500
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.