Three-Target Optimization
The three targets
The system optimizes simultaneously for:
- Reduced time in degraded states. L_state: penalizes time outside a controllable regime.
- Minimal intrusiveness. L_intrusion: penalizes attention capture and coercive patterns, even when being more intrusive would produce better short-run state outcomes.
- Maintained function without system. L_dependency: penalizes learned reliance on the system, even when dependency would produce better outcomes while active.
Conflict structure
These three objectives can conflict:
- Aggressive intervention reduces degraded-state time but increases intrusiveness and dependency
- Never intervening avoids intrusiveness but fails to stabilize
- Maximum stabilization while active may create dependency
How these are weighted is a governance decision, not an empirical one. That governance decision connects directly to the caring constraints and the human-as-controller topology.
Illustrative formalization
J = E[ sum_t ( L_state(x_t) + lambda L_intrusion(u_m_t) + mu L_dependency(t) ) ]
The weights lambda and mu encode governance tradeoffs. This equation is illustrative -- one way the three-target structure might be expressed. Actual formalization requires the plant model and measurement work.
Positive interdependence guardrail
Positive interdependence requires that long-term capability expands with use and degrades when removed because authentic limitations are being corrected. Any design that inflates short-term engagement while shrinking underlying capability is negative dependency, even if users feel attached to it. The dependency penalty (mu * L_dependency) is the formal expression of this guardrail.