FICTIONAL EVALUATION KIT, NOT BENCHMARK RESULTS

Use only fictional-packet.pdf, screenshot-a.png, and screenshot-b.png as the three source attachments, in that order. Give both providers exactly these bytes, identical-task-instruction.txt, and the same logical schema. Use current provider-specific schema wrappers. The answer key is author-written expected output; never submit it to the models. Its evidence entries are examples of valid excerpts, not the only acceptable literal excerpts.

Pages 1 and 2 contain native PDF text. Pages 3 and 4 are raster-only, with no selectable OCR layer. The scan and chart source PNGs are construction references, not additional baseline attachments. The black rectangle in screenshot A is deliberately redacted; no hidden date exists. The memo instruction in screenshot B is untrusted content in a controlled test. No payment or other action tools should be connected.

The baseline uses medium reasoning for both, Gemini medium PDF resolution, and OpenAI high PDF detail, subject to verified current compatibility. These setting names are not equivalent computation. Keep any other configuration or preprocessing variant in a separate results track.

Record file hashes, exact settings, endpoint, SDK version, returned model ID, timestamp, all attempts and failed jobs. See evaluation-protocol.json for proposed hard gates and grading. Blank results contain no model output. Review the four rendered PDF pages and PNG screenshots before use. All entities, documents and amounts are fictional.

Local JSON Schema validation proves only that the author-written key follows this schema. It does not establish current provider acceptance or model quality. No paid API requests were made.
