On ResearcherBench OpenAI Deep Research has 0.84 faithfulness and 0.34 groundedness, Grok3 DeeperSearch 0.80 and 0.31, while Sonar Reasoning Pro has the lowest faithfulness (0.62) and the highest groundedness (0.68); on DeepResearch Bench the number of supported citations per task runs from 4.35 to 111.21. A support rate is computed over the claims a system chose to cite and says nothing about the claims it left uncited or about how many citations it gives.

The faithfulness anchor contributes the rate among cited claims, the groundedness anchor the share of claims that are cited at all, in the same table for the same systems, which is what allows the statement that they move independently. The volume anchor contributes that accuracy and count separate as well: 94.04 accuracy with 9.78 effective citations for one system, 81.44 with 111.21 for another. The ratio 111.21 to 4.35 is computed here, not printed in the paper.

Evidence that uncited claims in these reports are common knowledge needing no citation, which would make low groundedness harmless; or a benchmark where groundedness and faithfulness rise together across systems. Query: human classification of uncited claims in ResearcherBench outputs as citation-worthy or not.

A tool with 85 percent citation accuracy is 85 percent reliable. The 85 percent covers only the third to two thirds of claims that carry a citation.