RESEARCHInvestigateNEXT 12 MONTHS
The Off-Support Barrier: Why Semantic Safety Constraints Are Not Learning-Problem Invariants, and What Follows for Prior Design, Containment, and Verification
arXiv cs.LG — Machine Learning
Factual evidence
What the source reports
Paper uses singular learning theory to prove semantic safety constraints lie off-support, meaning data training alone cannot guarantee safety.
Open sourceOneBench interpretation
Institutional assessment
So what
Theoretical proof that data training cannot guarantee semantic safety reinforces model risk requirements for deterministic, out-of-band sandboxing of autonomous financial agents.
Do what
Brief your model risk team to require structural deterministic containment rather than statistical guardrails for agentic deployments.