Sheaf theory helps AI detect when its scientific models break

In a new paper on arXiv, researchers David Olivieri and Roque Hernández introduce a sheaf-theoretic framework for detecting scientific theory shift in AI agents. The key idea: instead of simply fitting equations to data, an artificial scientist must know when its existing representational language stops being transportable into a new regime. The authors model contexts as local-to-global structures using source, overlap, target, and validation charts. Failure of coherence is measured through five obstruction metrics: residual fit, overlap incompatibility, constraint violation, limiting-relation failure, and representational cost. The framework ranks candidate shifts by obstruction level, allowing the agent to decide whether to deform within its current language or extend it.

The authors evaluate on a controlled transition-card benchmark designed to separate simple deformation from true language extension. The main result: the intended deformation or extension is almost always the lowest-obstruction candidate, and transition types cleanly separate in the benchmark. A secondary constellation kernel provides representational similarity probing. The work deliberately avoids reconstructing historical paradigm shifts or solving open-ended theory invention. Instead, it isolates a finite, practical diagnostic: when does representational transport fail, making extension the coherent next move? For AI safety and scientific discovery, this offers a principled way to detect when models are becoming stale and need new conceptual tools.