A new task, not a new model benchmark

From reviewed answers to four expense rows

Six accepted task responses do not mean six exportable expenses. Two correctly ask for missing information; only four reach the CSV after separate approval.

All receipts and twelve A/B responses here are fictional and agent-authored. First-pass errors were deliberately constructed, not sampled from a model. The same agent returned criterion judgments in another browser context. That is workflow dogfooding, not an independent reviewer, measured prompt improvement, OCR evaluation, real purchase or tax assessment. Six other pack cases remain unrun.

1. Record the task, not just a winning score

The unrun twelve-case pack has exact expected fields, a structure contract and three manual criteria. A browser imported the authored response worksheet after preview and confirmation, recording six A/B pairs. R01 A returned the correct fields but cited only Change EUR 5.60. It passed the automatic checks and the companion's exact-quote check; the quote was still insufficient support for the paid total.

Recorded project before review · Original outbound review packet

2. Return judgments without replacing the evidence

The local recipient editor returned all twelve criterion assessments with an explicitly same-agent label. R01 A failed evidence quality. The other A examples guessed USD, added JPY decimals, rounded KWD, obeyed an embedded footer or computed a missing final total. All six A responses were rejected. Six authored B revisions were accepted under this task policy; two were correct clarification responses with null extracted fields.

The sender previewed and confirmed the return. Exact prompts, outputs, model/authorship labels and recording dates remained unchanged. The recorded dates and review-start dates are browser-device dates, not execution receipts. These files are unsigned and not blinded.

Returned judgments · Applied project · Offline all-recordings evidence HTML · Comparison report

3. Review downstream meaning separately

The six B source/answer pairs were copied verbatim from the applied project into the ledger. The companion's downloaded R01 extraction prompt exactly matched the recorded B prompt. The tool did not infer expenses from accepted reviews: each ready row required a new checkbox approval.

There is no combined monetary total across currencies. R09's merchant is literally =SUM(1,2) Studio. The expense CSV prefixes an apostrophe and shows that change before export; the work file retains the original source and answer. This is not guaranteed protection for every spreadsheet importer.

Six-row portable receipt work · Actual four-row browser CSV. Opening the work file restores no approvals. In a fresh browser context, previewing and approving the same four rows reproduced the identical CSV. A fifth matching row triggered the duplicate warning; it was separately exercised and excluded from this published CSV, never silently removed.

4. Read what the import actually established

Python's standard CSV reader parsed the actual browser download into four five-field rows. A parameterized in-memory SQLite table with TEXT columns retained every cell, including the merchant's literal apostrophe and exact amount strings. Decimal arithmetic reproduced the four separate totals without binary floating-point rounding. Inspect the independent-parser receipt.

No LibreOffice, Excel, Google Sheets, accounting-provider import or real payment was performed. Spreadsheet/date/number interpretation remains untested; use text columns and inspect values. The local SQLite exercise does not establish compatibility with your accounting workflow.

Keep the evidence and know its boundaries

Artifact sources and SHA-256 hashes identify the frozen browser downloads, not authorship, model execution, payment or an authenticated review. Recorded project and returned packet use existing Test Bench formats; receipt work is separate. No automated project-to-ledger transfer or source authenticity verification exists.

You may reuse or adapt these fictional sources, prompts and authored examples, including commercially, without attribution or warranty. This grants no rights in receipts you supply.