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.
- Greedy50% rules
- Sample 167% rules
- Sample 250% rules
- Sample 350% rules
- Sample 475% rulespicked
Extracted fields
| Item | Qty | Unit price | Amount | Remove |
|---|---|---|---|---|
Accounting checks
goes to human review
- Line 1: qty × unit price = amountexpected 120,000 · found 60,000
- Items add up to the subtotalexpected 60,000 · found 60,000
- Subtotal − discount + service = total (tax already included)expected 60,000 · found 60,000
- Card charge = totalexpected 60,000 · found 60,000
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.