Work in this area focuses on AI systems that move beyond fixed-objective optimisation toward architectures capable of operating on evaluation itself. This includes epistemic generativity, evaluative system design, and structural approaches to out-of-distribution failure and objective misspecification.
This paper argues that out-of-distribution failure is not merely degraded performance but a structural signal of epistemic enclosure: current machine learning systems cannot treat failure as evidence against their own evaluative standards, only as deviation within them. details
Until now, no explicitly epistemic theory of generativity has been formulated. This paper defines the structural threshold between optimisation and generativity, locating agency and responsibility in control over evaluative criteria rather than behaviour. details