OfficeQA Pro
Databricks / Professional work · Retrieves information, calculates, and answers from large collections of real financial documents.
What it measures and how
Fixes Pro's 133 tasks, separate from Full's 246 and Pro V2's 90. End-to-end retrieval and directly supplied answer pages differ. Claude Agent SDK, Codex SDK, and Antigravity CLI results belong to their systems, not a common base-model evaluation. Measures document understanding, not output-file quality.
How this evidence is used
Under review: official updates continue, but results mainly compare product systems. Raw aggregation needs checking. Retained as document-understanding reference; different tool results are not attributed directly to base models.
Evaluation limits & attribution
Code and task licenses do not automatically cover external results. Old/new sets, end-to-end retrieval, and supplied answer pages stay separate. Results are not automatically collected.
Data license: Code is Apache 2.0; task data is CC BY-SA 4.0 and has moved to a Hugging Face dataset requiring approved access.
Results are published by Databricks. Original scores and the NexGateHub consensus index use different scales and cannot be added directly.