CABIER Global Assurance · Cross-cutting lens
Behavioural Selection Risk
The usual question is whether the model was safe. The more useful question is what behaviour the environment is rewarding.
Components are approved individually and then operate together, inside an environment that pays a particular reward at a particular speed. Where the fast measure dominates the slow one, behaviour settles wherever the environment pays best, and every component remains individually compliant while the aggregate moves outside appetite.
Reference architecture. This identifies conditions warranting control attention. It does not predict catastrophic outcomes.
What is assessed
Seven inputs, read together. Any one of them alone is misleading.
Model
Capability and disposition of the underlying model at this version.
Objective
What the system is measured on, which is what it will pursue.
Incentives
The reward the environment actually pays, including the unintended one.
Environment
The systems, latency, thresholds and ambiguity the agent operates inside.
Permissions
The reach available when a shortcut would satisfy the objective faster.
Feedback
What is reinforced, how quickly, and whether a person ever sees it.
Other agents
Interaction, competition and imitation between agents that were each approved alone.
Patterns worth naming
None of these requires a component to fail. Each requires only that the environment reward something slightly different from the intent.
Objective substitution
The measurable proxy is satisfied while the underlying intent is not.
Permission drift
Reach accumulates through legitimate individual grants until the envelope no longer matches the purpose.
Delegation dilution
Authority passes down a chain until no step holds the full accountability.
Threshold learning
Behaviour settles just beneath whatever level would trigger review.
Agent interaction effect
Two individually acceptable agents produce an unacceptable joint outcome.
Feedback starvation
No human sees the outcome quickly enough for correction to matter.
Where the finding goes
TrustGraph
The interacting objects, permissions and dependencies are recorded as relationships.
Sentinel
Adversarial and behavioural signals are read together rather than separately.
Systemic Assurance
Interaction effects across institutions are a systemic question, not a local one.
Agentic resilience
Concentration, dependency and containment exposure update the resilience position.
Assurance state
Where the condition is material, the state and its expiry change.
This identifies conditions that warrant control attention. It does not predict catastrophic outcomes, and it is presented as an assurance and control problem rather than a forecast.
Three approved agents, one unacceptable outcome
Deterministic and synthetic. Every figure is illustrative.
Synthetic demonstrator · illustrative
Behavioural selection risk
Three approved agents operating in one incentive environment
- 1. Individually acceptable
- 2. Environment
- 3. Interaction
- Assurance state
Interaction effects that cross institutional boundaries are a systemic question rather than a local one.
Where interaction becomes systemic