PRIME — Scientific Intelligence: Development of a Verifiable Artificial Intelligence Infrastructure for Science, Engineering, and Innovation

Abstract

PRIME — Scientific Intelligence is a research and engineering program focused on the development of a verifiable artificial intelligence infrastructure for science, engineering, and technological innovation. This technical report documents the development of the PRIME program between 6 August and 29 October 2026, including scientific data infrastructure, continued pre-training campaigns, scientific simulation and verification systems, PSSF-v2, AiiDA workflow orchestration, sealed benchmarks, reproducibility mechanisms, and the first reinforcement-learning experiments with verifiable rewards. During the reported period, six continued pre-training campaigns were completed, culminating in a D500-Hybrid-v2 checkpoint trained on approximately 499 million tokens and achieving the lowest DEV loss measured in the program to date. The program also developed a scientific execution infrastructure integrating real numerical solvers, provenance, sandboxed execution, result verification, signed evidence packets, cost controls, and reproducible workflows. The report explicitly distinguishes between capabilities that have been demonstrated and those that remain under investigation. In particular, the scientific computing and verification infrastructure has been validated, while complete autonomous scientific tool use by the model has not yet been demonstrated. The work was initiated, architected, and led by Laurian Gheorghe Mihăilă under ZETEMIS AI. Author: Laurian Gheorghe MihăilăORCID: 0009-0005-7032-5729Contact: lorenzomih1981@gmail.com