Evident Machines

Much of scientific discovery in electrical engineering happens through physical experiments, where real-world measurements reveal things a simulation cannot capture. This research is inherently iterative, moving back and forth between theory, computation, and the physical lab bench. As frontier AI models are increasingly used in this kind of work, there is no clear way to measure whether they can exercise real scientific judgment: designing a research protocol, investigating an unexpected result, or weighing competing physical constraints. Without that measurement, it is unclear how much of this research can actually be entrusted to AI models.

Evident Machines builds benchmarks around long-running electrical engineering research and engineering projects rather than isolated test questions. The tasks combine specialized software, real experimental measurements from the lab bench, and physical constraints, mirroring the iterative theory-to-bench workflow of real EE research. By making model failures specific and reproducible, the benchmarks aim to give a concrete target for improving models and a way to check whether those improvements actually hold up in physical lab research. The company is building these evaluations together with scientists and engineers who contribute tasks drawn from work they know deeply.

The site frames its work around frontier models newly opening possibilities for electrical engineering research, making it urgent to measure their scientific judgment now.