- Selected model attempts
- 3
- Selected model charges
- $0.0000 USD
- Human editing
- Not recorded
kingy tested
Original accepted output
{
"invoice_id": "K-1042",
"total_usd": 127.50,
"due_date": "2026-09-20",
"bank_account": null
}Qwen3-4B-Q4_K_M.gguf · seed 4101. Retained output for the original input. Changing the recipe does not generate or verify a new result.
Reproduce the outcome
- Download and extract the public evidence ZIP.
- Run python3 verify_scores.py to recompute the original scores without inference, network requests or downloads.
- Inspect raw/local-4B-L01-1.json for the exact request and result.
- To make a new invoice variant, copy the adapted prompt and run it only in your chosen local environment.
- Compare every field and type with your source invoice; preserve null for absent facts. Record all failures as attempts.
Original input and exact settings
- Synthetic invoice K-1042, supplied in the exact prompt
- temperature
0.6- top_p
0.95- max_tokens
256- context
4096- seeds
[4101,4102,4103]
Original prompt
Return JSON only with invoice_id, total_usd (number), due_date (ISO date), and bank_account (null if absent). Invoice K-1042. Total USD 127.50. Due 2026-09-20. Purchase order PO-88. Do not infer missing facts. /no_think
Attribution and rights: Kingy synthetic task and actual redacted output evidence; model weights remain under their separate license and are not redistributed.
Sources & limits
Small Qwen3 models: retained JSON pilot
Record checked 2026-09-07.
Provider charge was $0. Electricity, hardware and active repair time are unknown.
Only this synthetic invoice was observed. Adapted inputs are untested.
Measured success and timings apply to the exact model/runtime/settings and one M4 Pro.