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.
- Greedy100% rulespicked
- Sample 1100% rules
- Sample 2100% rules
- Sample 3100% rules
- Sample 4100% rules
Extracted fields
| Item | Qty | Unit price | Amount | Remove |
|---|---|---|---|---|
Accounting checks
would be accepted
- Line 1: qty × unit price = amountexpected 16,363 · found 16,363
- Items add up to the subtotalexpected 16,363 · found 16,363
- Subtotal − discount + tax + service = totalexpected 17,999 · found 17,999
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.