When Generation and Verification Both Scale: A Capability–Reliability Frontier for AI-Mediated Scientific Publishing

Generative artificial intelligence (AI) may increase manuscript arrivals and screening capacity. We model review-cycle arrivals at the deduplicated manuscript-series level in an AI-first network with direct disposition, randomized limited audit, full specialist review, and accountable editorial decisions. Station-specific queueing constraints couple to route-conditioned false-passage and false-rejection probabilities, class-conditional attacks, generator–reviewer dependence, and common-shock AI-panel errors. We derive a closed-form interior minimum-audit policy, conditions requiring corrective human assessment under editorial ratification, and an optimized capability–reliability frontier. With reliability, feedback, staffing, and policy-independent utilization limits fixed, an exact linear program solves the joint limited-audit and specialist-route allocation. Deterministic optimization, 20,000 independent cohort-accounting replications, five finite-manuscript and stationary-queue experiments, and 16 independently scrambled Sobol designs check the results. In the single-reviewer limited-audit slice, the maximum safe arrival multiplier decreases from 1.309 to 1.001 as the generator–reviewer stress index ξ_GA rises from 0 to 0.5. Allowing the separately provisioned specialist route attains the AI/editor ceiling 1.70 at three stress settings; this is not an equal-resource comparison. The analytical mean route time 3.8515 lies within the Monte Carlo 95% interval [3.8488,3.8581]. Only 14.58% of the declared sensitivity domain is feasible. AI review expands safe capacity only when compute, editorial, audit, and field-matched specialist resources jointly satisfy delay and error constraints. When common-mode AI error exceeds tolerance, a sufficient net correction is required; statistical independence is not. This preprint record provides the revision-41 main manuscript (AI_research_41.pdf), its supplementary material (ESM_1.pdf), and the corresponding source and PDF bundle (AI_research_41_verified_bundle.zip). The bundle contains the main and supplementary LaTeX sources, compiled PDFs, figure PDFs, auxiliary files for cross-document references, and a README. The associated reproducibility materials remain available in the earlier, unchanged Zenodo version 1.2.0 archive at https://doi.org/10.5281/zenodo.22826958. That archive contains the revision-37 manuscript source together with the archived computational materials. The revision-41 main and supplementary sources provided here are distributed separately from that earlier archive. This record is a preprint and source release, not a new computation release, and does not claim a new execution of the archived numerical experiments.