Safety and Risk Surfaces under Projection

Structural Problem

Safety boundaries and risk assessments for advanced AI systems are typically defined in a specific analytical frame — a particular set of metrics, test conditions, and evaluation criteria. The structural problem is that these boundaries are not stable under projection: when the system is observed or operates in a different frame (different context, scale, or deployment condition), the safety surfaces shift in ways that invalidate the original assessment.

This is not a testing coverage problem. It is a structural property of systems with emergent behavior: the risk surface itself changes shape depending on the projection through which it is observed. A system that appears safe under one set of evaluation criteria may exhibit entirely different risk characteristics when deployed in a context that projects onto a different region of the stability space.

System Context

This application addresses advanced AI systems where capability, safety, and risk are interrelated in non-linear ways. The relevant system boundary includes model behavior, deployment environments, evaluation frameworks, and the interaction between capability development and safety constraints.

The structural challenge is particularly acute for systems approaching or crossing capability thresholds where emergent properties alter the risk landscape. Safety assessments performed at one capability level may not transfer to the next, creating a moving target that static risk frameworks cannot capture.

Diagnostic Capability

This application provides structural analysis of safety and risk surfaces across projection frames, identifying conditions under which safety boundaries shift or collapse. The diagnostic output maps stability classes — regions of the capability-safety space where risk properties are structurally stable — and identifies failure modes associated with transitions between classes.

• Risk surface mapping across multiple projection frames
• Stability class identification for safety boundary persistence
• Failure mode analysis at stability class transitions
• Projection-dependent safety assessment framework

Typical Failure Modes

• Projection collapse where a safety boundary that holds under evaluation conditions disappears under deployment projection
• Emergent risk shift where capability development moves the system into a region of the stability space where previously assessed safety properties no longer apply
• Assessment frame lock-in where safety evaluation is conducted in a fixed frame that does not capture the structurally relevant risk dimensions

Example Use Cases

• Pre-deployment safety projection: Structural analysis of whether safety properties established during evaluation persist under target deployment conditions
• Capability threshold risk assessment: Identifying stability class boundaries before planned capability improvements to anticipate safety surface shifts
• Multi-frame safety validation: Assessment of safety boundaries across multiple projection frames to identify the most structurally constraining conditions

Strategic Relevance

As AI systems grow in capability, the structural relationship between capability and safety becomes the determining factor for responsible deployment. Static safety assessments that do not account for projection effects provide false assurance. This application enables structurally grounded safety analysis that remains valid across deployment conditions and capability levels.

SORT Structural Lens

The SORT framework addresses this application through four structural dimensions, each providing a distinct analytical layer.