Audit Finds Measurement Gaps in Frontier AI Forecasting

An audit released on August 12, 2026, evaluated how historical records influence predictions regarding advanced artificial intelligence. The analysis, identified by the code 2608.14903, focused on 62 AI systems and included 12 benchmarks, seven assessment criteria, 144 evaluated events, and 27 source documents. Key findings revealed significant deficiencies, with only seven systems possessing both compute estimates and a METR 50 percent task horizon observation. The report uncovered that 19 of the 27 closed systems lacked training compute data, particularly all 2026 releases, while none of the 35 open-weight systems met METR horizon observation requirements, prompting concerns over forecasting accuracy.