Bayesian Inference and Decision Audits for Public Archives of Frontier AI Evaluations

Public leaderboards for artificial intelligence systems are widely treated as definitive rankings, but a new analysis argues they are actually selective time series distorted by reporting conventions and missing data, making terminal comparisons unreliable [1]. The study, posted to arXiv on June 15, 2026, examines repeated public archives from LiveBench and the Open LLM Leaderboard v2 as its primary longitudinal record, with LMArena used as a preference stress test and GAIA and tau-bench contributing limited agentic pilots [1]. Language model benchmarks are standardized tests designed to evaluate performance on tasks such as language understanding, generation, and reasoning, and are maintained by academic institutions, research organizations, and industry players to track progress in the field [6]. The authors frame the evaluation record as a Bayesian inference problem. Under a fixed reporting convention, a constructed terminal-only example covering 1,000 systems is shown to be compatible with two distinct pre-terminal histories, yielding times of 23.03 or 75.13 to reach within 0.05 of the performance ceiling under the same terminal-tail model [1]. This finding underscores how the same final snapshot can mask fundamentally different development trajectories. The paper introduces a candidate selection-aware frontier model and tests it across four diagnostics: synthetic recovery, objective-archive prediction, preference transfer, and uncertainty calibration. The model fails all four, and fixed audit gates reject its stronger claims [1]. Explainable AI research has long sought methods to provide humans with intellectual oversight over algorithms, countering the "black box" tendency of machine learning where even designers cannot explain why a system arrived at a specific decision [5]. The audit protocol described in the paper operates in a similar spirit, reconstructing public evaluation histories to isolate a verified timing boundary and falsify unsupported frontier claims [1]. Generative AI, a subfield that uses models to produce text, images, video, and other data, has seen a significant increase in prevalence since the AI boom of the 2020s, driven by improvements in deep neural networks and large language models based on the transformer architecture [4]. The field of artificial intelligence itself was founded as an academic discipline in 1956 and has gone through multiple cycles of optimism followed by periods of disappointment known as AI winters, before funding and interest surged after 2012 with the rise of deep learning [3]. The new archive-and-adjudication protocol offers a method for scrutinizing the evaluation claims that accompany this rapid progress, treating public benchmarks not as final scoreboards but as incomplete observational records requiring statistical reconstruction [1].