Evaluating AI-Generated Scientific Questions with Discovery Plane Theory: A Repeated-Measures Computational Benchmark Across Twenty Scientific Fields

Scientific artificial-intelligence benchmarks predominantly evaluate answers, reasoning, code, hypotheses, or workflow completion after a task has been specified. This study evaluates an earlier epistemic object: the scientific question itself. A reproducible computational benchmark operationalised Discovery Plane Theory (DPT) to analyse 1,000 AI-generated questions derived from 100 seed problems across 20 scientific fields. Each seed was transformed by ten controlled reframing strategies, creating a repeated-measures design. Six automated proxy components operationalised Discovery Complexity (DC), and four operationalised Question Compression (QC). Under a prespecified exploratory threshold of .65 on both normalised coordinates, 541 questions (54.1%) were classified as prospective Region IV and six were identified as Discovery Frontier candidates. Repeated-measures analyses showed strong strategy effects for DC, χ²₍F₎(9) = 853.43, p < .001, Kendall's W = .948, and QC, χ²₍F₎(9) = 898.52, p < .001, W = .998. Cochran's Q(9) = 391.63, p < .001, confirmed strategy-dependent Region IV membership. Cluster-bootstrap estimates linked DC to novelty (ρ = .323, 95% CI [.277, .371]) but did not support an association between QC and usefulness (ρ = −.032, 95% CI [−.096, .030]). A top-quartile novelty-plus-usefulness baseline overlapped only modestly with Region IV (Jaccard index = .214). Threshold and component-weight sensitivity analyses showed that absolute Region IV prevalence depended on the chosen operational conventions, whereas strategy ordering was moderately stable. The results demonstrate that Discovery Plane Theory can be operationalised as a reproducible benchmark for analysing AI-generated scientific questions and provides a structurally distinct evaluation framework relative to conventional novelty–usefulness screening.