Abstract
<p>Grant peer review increasingly depends on fine-grained numerical scores to rank many formally excellent proposals. In the era of generative and agentic AI, this architecture may be losing resolution. We analyse score-distribution data from EU Marie Skłodowska-Curie Actions (MSCA) Postdoctoral Fellowships and show sharp compression of evaluation scores in the 2025 call. Compared to 2024, there was a 64.6% increase in the volume of applications received, low-scoring proposals became rare, and ~20% of proposals reached a score of ≥95%. The timing is consistent with AI-assisted optimization, but alternative structural explanations remain possible. Nonetheless, the score compression exposes a structural vulnerability of highly standardized evaluation systems: evaluation precision near the top of the scale may be more apparent than real. We use this pattern to derive a cross-funder diagnostic framework: compression should be strongest in schemes with explicit scoring rubrics, become more visible after broad diffusion of advanced AI drafting and reasoning tools, and intensify as these tools gain capabilities. Data from Human Frontier Science Program (HFSP) Postdoctoral Fellowships and Research Grants, which first screen short Letters of Intent before inviting selected applicants to submit Full Proposals, show broadly similar score distributions across years, with no comparable compression. Cross-funder score tracking can therefore distinguish scheme-specific disruptions from broader temporal shifts and identify designs that preserve evaluative resolution. If compression is becoming more widespread, funders should reconsider whether fine-grained ranking remains an efficient and fair way to select among many excellent proposals.</p>