Affective State as Control Variable in Adaptive Governance Systems

This paper formalizes affective state as a control variable in adaptive governance systems. It argues that the standard state-only policy formulation π (a∣s) (a|s) π (a∣s) is insufficient for systems with internal affective dynamics, because identical external states can produce sharply different action distributions depending on the operative affective state. The paper introduces the affective state vector e (t) ∈Rke (t) ᵏe (t) ∈Rk as a low-dimensional control state derived from internal dynamics through an affective readout e (t) =Pz (t) e (t) =Pz (t) e (t) =Pz (t). This leads to the emotion-conditioned policy formulation π (a∣s, e) (a|s, e) π (a∣s, e), where action selection depends jointly on external state and internal affective activation. Alignment failure is modeled as an affect-conditioned regime transition in policy space, with failure probability F (t) =Pr (misaligned action∣st, et, Λt) F (t) =Pr (misaligned action sₜ, eₜ, ₜ) F (t) =Pr (misaligned action∣st, et, Λt). The paper develops regime-dependent effects for affective states such as desperation, calm, loving, and anger, and connects these to failure modes including coercion, reward hacking, sycophancy, and harsh directness. The paper also establishes cross-theory couplings between affective state and the broader DFG framework: affect acts as a VST instability mode selector, an RBIT resolution-pressure observable, and a NAT coordination bias field. Controlled mechanism tests provide support for the associated AGM predictions, while direct behavioral validation of the policy-conditioning claims remains a target for steering-accessible datasets.