Summary

DyFram measures AI risk as a product of two inseparable elements: Harm Potential and Contextual Activation. In compact form:

Risk = Harm Potential × Contextual Activation

This formulation captures both what an AI system can do by design and what happens when it is put into a specific use environment. The resulting R score is auditable versioned and designed for operational use by governments and large institutions.

What DyFram means by Harm Potential

Harm Potential is the danger inherent to the system itself. It is determined by core MDCM fields such as the Capability Profile and the Impact Profile. Examples of high Harm Potential include systems that can autonomously make or materially influence life changing decisions infer sensitive attributes without consent or operate with high‑precision physical effectors.

What DyFram means by Contextual Activation

Contextual Activation describes how the system is used and who controls it. It covers Use Case Definition Control Mode Actor Classification and Environment. A model with modest Harm Potential may become high risk when deployed in an unsupervised emergency setting or used by untrained operators; conversely a powerful model may be managed safely when placed behind strict human oversight and limited access.

Why the multiplicative formulation is practical

• Reflects real causality Harm exists in the system but risk emerges when the harm is activated by context. Multiplication matches that causal intuition: if either factor is zero the realized risk is negligible.
• Supports proportionate governance By separating intrinsic danger from activation, DyFram guides proportionate obligations. High Harm Potential with low activation may need strong organizational controls; low Harm Potential with high activation may require strict user gating.
• Enables auditable decisioning Both Harm Potential and Contextual Activation fields are versioned and recorded. The R score becomes reconstructable during audits and reviews.
• Preserves sovereign tuning Governments can tune thresholds that influence how Contextual Activation maps to governance G levels while the core meaning of those levels remains standardized across jurisdictions.

Concrete examples

Analogy 1. A tiger is dangerous by nature. In a forest the hazard is present but contained. Put that tiger into a crowded market and the activation multiplies the danger. The combined effect is what matters for governance.

Analogy 2. A parked car has inherent danger. On a busy highway at 80 mph the activation by context makes it an imminent public safety risk. Similarly an AI model that can influence safety outcomes becomes far riskier when placed in high speed or high stakes operational contexts.

AI example. A powerful generative model has high Harm Potential for misinformation. If used only in closed research with expert review the Contextual Activation remains low. If the same model is embedded in an automated public communications pipeline with no human review the activation is high and the R score rises accordingly.

Operational consequences for governance

• DyFram R scores instruct which governance layers apply from UGC to CSG UGL and OGL and whether CPP red band measures are needed.
• R scores inform Minimum Capability Requirement settings, deciding when operator certification or supervision is mandatory.
• They drive clear audit trails showing why a specific obligation applied which actor owned a duty and what JAL package influenced the outcome.

Why this definition is more accurate than single label approaches

Many frameworks label systems as high risk based on capability alone or on sector labels that blur real differences. DyFram’s definition is practically accurate because it mirrors how harms actually occur. Risk is not an abstract property of a model; it is the realized exposure when model capabilities meet context. This avoids both false positives and false negatives in governance while making obligations defensible under audit and legal scrutiny.

Closing

DyFram’s Risk = Harm Potential × Contextual Activation is simple but powerful. It turns classification into actionable governance by making risk measurable explainable and tied to both system identity and deployment reality. For governments and institutions that must govern under law and public scrutiny this is not a theoretical improvement. It is an operational necessity.