# Attention and review

## 3. Results

Across all three cohorts, retrieval-augmented review reduced median time-to-first-comment from 4.2 days to 1.6 days.
The effect was consistent across the 2023 and 2024 submission windows (Table 2).
Reviewers who received sentence-level provenance markers flagged 31% more unsupported claims than the control group.
We found no significant difference in overall recommendation scores between conditions (p = 0.41).
The improvement holds for submissions under eight pages, where reviewers read the full text.

## 4. Discussion

These results suggest that the bottleneck in peer review is orientation rather than judgement.
Reviewers do not lack the expertise to evaluate a claim; they lack a cheap way to locate the claim's support.
Our intervention is therefore best understood as a navigation aid, not an evaluation aid.
The 31% increase in flagged claims is obviously the headline result.
We caution against reading the null result on recommendation scores as evidence of no effect.
Power to detect a difference smaller than 0.4 points was low in all three cohorts.