DataSpace measures data-agent integration across heterogeneous evidence
Its 66.34% best score and 15.36-point fixed-model harness spread are benchmark results, not a predicted production ceiling.
DataSpace contains 410 cross-language tasks and 7439 artifacts totaling 15.01 GB across CSV, JSON, SQLite, Markdown, PDF and video. Agents receive a question and workspace and must return the complete requested table. A deterministic evaluator handles alignment, type/precision normalization and row ordering.
Across six tested multimodal models and five harnesses, the best accuracy is 66.34%. With the backbone fixed, harness choice produces a 15.36-percentage-point spread. The paper identifies multimodal evidence integration and joins as recurring difficulties across these backbones.
Those findings show that harness choice matters in this evaluation; they do not prove any arbitrary production workload must score below 66.34%. A production task distribution may be easier, harder or simply different.
For a data-agent trial, preserve the complete expected table and compare harnesses on the same model and workspace. Diagnose retrieval, joins and output normalization separately before attributing every miss to the model.
Editorial correction (2026-10-13 UTC): Removed the guaranteed production decline while retaining the integration experiment.