Do Headline Claims Match the Evidence? A Systematic Audit of Peer-Reviewed Journal Evaluations of LLM Serving

Background: Peer-reviewed journal evaluations of large language model (LLM) serving operations include claims regarding latency, throughput, SLO achievement, efficiency, robustness, and serving behavior. Aim: To evaluate whether selected headline claims meet prespecified minimum claim-specific evidence requirements while simultaneously assessing external validity, stress testing, comparability, and reproducibility. Methods: We systematically reviewed peer-reviewed journal articles identified through Scopus, Web of Science Core Collection, and IEEE Xplore through 25 September 2026. One primary operational claim was selected per distinct study and assessed against a prespecified claim-class evidence matrix. Screening and coding were conducted by the sole author with AI-assisted support; no independent second-coder reliability statistic was calculated because independent recoding was not completed. Results: Seventy studies were included. Latency (21/70, 30.0%) and throughput/goodput (20/70, 28.6%) were the most common primary claim classes. Sixty-nine claims (98.6%) met the prespecified minimum evidence criteria and one (1.4%) partially met them. A coded adverse-condition experiment was present in 65/70 studies (92.9%), but this was dominated by load/resource stress; explicit component failure appeared in only 5/70 (7.1%). Only 25/70 studies (35.7%) exposed a study-specific reproducibility artifact and 12/70 (17.1%) explicitly reported repeated runs. Generated traffic dominated (51/70, 72.9%); no study used live production traffic. Conclusion: Minimum claim-specific evidentiary adequacy was common in this journal-defined corpus, whereas reproducibility and production-proximate evidence were substantially weaker. The near-degenerate adequacy result should be interpreted cautiously because independent reliability coding was not completed.