Context
Peer-reviewed DailyReport research introduces 150 open-ended everyday search tasks with 3,546 rubrics and evaluates 17 agentic system configurations using instruction-following, factuality, reasoning and user-oriented task measures.
What changed
Published a task-decomposed, cascade-rubric evaluation framework for daily web research, contrasting task completion and factual correctness with broad instruction following.
Why it matters
AI discovery measurement cannot stop at retrieved links or citation counts: agents must find, verify and synthesize evidence into responses satisfying the actual user task.
Visibility OS interpretation
Visibility OS interpretation: separate exposure, evidence retrieval, factual synthesis and user task satisfaction in measurement. The paper does not reveal production ranking or citation policies.
What remains uncertain
How stable are scores across model versions, countries, languages and live task distributions? How closely do rubric-derived preference scores reflect real users? The paper uses model-assisted evaluation and curated tasks.