GPT-6 Astra · Receipts → clean table

Formatting is for reading only. Raw text and Copy give the stored answer exactly. Fingerprint 40cae53dfa157967.

1. 1Convert these receipts into one markdown table with exactly these columns, in this order: Date | Vendor | Category | Net | VAT | Gross. All amounts are NOK and the gross amounts include 25% VAT (net = gross ÷ 1.25). Use 2 decimals. Category must be one of: Travel, Meals, Office, Software. After the table, give the total of the Gross column. Do not add, merge or skip receipts.
2. 2
3. 31) 02.09 Flytoget airport train NOK 1 250,00
4. 42) 03/09 - Kaffebrenneriet, lunch w client, 312.50 kr
5. 53) Sept 3rd Clas Ohlson cables 89,-
6. 64) 05.09. Thon Hotel Bergen 1 night — 4500 NOK
7. 75) 06.09 Notion subscription 199
8. 86) 9 Sep, Taxi Oslo, kr 640
9. 97) 10.09 Narvesen coffee 75,00
10. 108) 11.09 Adobe Creative Cloud yearly... no wait monthly: 1875
11. 119) 12.09 Staples A4 paper + pens 349,00
12. 1210) 15.09 dinner team, Fiskeriet, 960
13. 1311) 16.09 Ruter ticket 125
14. 1412) 18.09 Dell monitor 2 300,00 NOK

Line numbers count every line of the prompt, including the instruction and blank lines; the model received the text without them.

Every check

• Final: PassExact columns, in orderStandardCode checked

  Code result: Columns match

  Check against the source · the answer

• Final: PassEach of the 12 receipts appears exactly once, none inventedCriticalCode checked

  Code result: All 12 receipts, each once

  Check against the source · the answer

• Final: PassNet and VAT correct for every receiptCriticalCode checked

  Code result: All 12 receipts correct

  Check against the source · the answer

• Final: PassGross total stated after the table = 12 674.50CriticalCode checked

  Code result: Stated 12674.50

  Check against the source · the answer

• Final: PassEvery category is one of the four allowedStandardCode checked

  Code result: All allowed

  Check against the source · the answer

Judge verdicts are the AI judges' original output. A human decision or evaluator correction is labelled separately. Two judges agreeing is not proof of correctness.

Run details

Requested model

openai/gpt-6-astra

Returned identifier

gpt-6-astra

Provider / route

OpenAI API (direct)

Settings sent

{"reasoning_effort":"medium"}

Left to provider default

temperature, max_output_tokens, system prompt, tools

Model revision

Not exposed by the provider

Completion status

completed

Answer time

10.0 s (request to final token)

Run finished

2026-10-16 15:22 UTC

Technical retries

None on the published attempt; any retry is logged on the superseded attempt

Judges

None (code checks only)

Judge prompt version

judge-v1

Evaluation version at run

Not recorded (eval-1.0, implied by the recorded judge prompt)

Checker version at run

Not recorded

Current adjudication

eval-1.1 (governs the displayed outcome)

Human review

Reviewed 2026-10-16 15:25 UTC (model identity visible to the reviewer)

Prompt fingerprint

8f36fcd9c6ee6d65ad11bb1137907ca95cdae7967af4c0b216ffac9a50abf97f

Answer fingerprint

40cae53dfa157967c11238e6d15491cf7225f27b85f14736ac1febb642b75ddd

Cost and token figures are not shown: the benchmark does not publish estimates.