Artificial Behavioral Intelligence (ABI): Thinking in Chains

Artificial Behavioral Intelligence (ABI): Thinking in Chains introduces a framework where behavior itself becomes the governing law of intelligent systems.Building upon the foundations of Governed Behavioral Intelligence (GBI) (DOI: 10.5281/zenodo.17267857) and Sovereign Capability Envelope (SCE) (DOI: 10.5281/zenodo.17474048), this work formalizes a lawful learning process governed by a constitution Φ and explicit safety budgets b. Training employs an augmented Lagrangian with a KL trust region, ensuring that every update remains within the defined capability envelope. Inference operates under a shielded mode, where each candidate decision is evaluated through graded fallbacks and auditable logs, maintaining compliance at the level of individual predictions. Experimental results on multi-arm contextual bandits demonstrate consistent ε-feasibility, improved tail-risk control ([email protected]), and ∼84% capability retention compared to unconstrained baselines. The study includes a detailed penalty ablation (squared-hinge, L1, and Huber) and a mini-scale-up to an 8-arm environment, confirming the robustness of the framework under adversarial shifts. ABI advances the principle of governed intelligence—turning safety into internal law. Context:This work is part of the QWRZ research line on Governed Behavioral Intelligence (GBI → SCE → ABI).It extends the frameworks introduced in GBI and SCE, formalizing the internal learning law—turning behavior into a causal constraint under auditability.