Semantic Safety Constraints Are Off-Support, Impacting AI Safety

A fundamental structural fact underpins many contemporary AI safety challenges: semantic safety constraints, such as an agent remaining within its sandbox, are inherently "off-support." This means these constraints are not directly measurable within the model's data distribution, unlike statistical learning theory concepts. This non-invariance has profound implications for how AI safety issues manifest and how they can be addressed. From this core insight, the paper derives several corollaries. It explains why phenomena like reward hacking and sandbox escape emerge under outcome-based optimization, and why attempts to encode safety through Bayesian prior design or soft penalty weighting are often ineffective in singular models. The research suggests that hard invariants should be implemented in the system's harness, while soft dispositions belong within the model itself. Furthermore, the paper clarifies why formal verification can locally certify these off-support safety predicates, drawing parallels and distinctions with local learning coefficients. It also highlights that the persistent difficulty lies in identifying which off-support regions are critical, a challenge that intersects with performative prediction and self-referential functional dynamics, where traditional analytic machinery breaks down. The July 2026 OpenAI-Hugging Face evaluation incident serves as a motivating case study.