From Orchestration to Integrated Judgment Optimization : A Boundary-First Architecture for Dynamic Teams, Version-Governed Structural Autonomy, and Human-Accountable AI

Most agent-system research proceeds bottom-up from models, tools, memory, role assignment, and orchestration. This paper develops a complementary boundary-first architecture motivated by the public LimFlex wrapper perspective. Boundary-First Integrated Judgment Optimization (BF-IJO) begins with the conditions under which a candidate may become an executable judgment and derives the required processing path, judging configuration, evidence acquisition, authority, human escalation, external commitment, and subsequent update controls. It separates hard admissibility constraints, judgment-validity conditions, and bounded performance optimization. Orchestration, Dynamic Variable Teams, and Version-Governed Structural Autonomy are recast as processing-path, judging-configuration, and judgment/update-lifecycle control surfaces of the upper-level problem. The paper synthesizes recent evidence on adaptive orchestration, system-level prompt optimization, calibrated act-versus-escalate decisions, evidence-decision decoupling, subagent privilege management, governed capability evolution, and semantic rollback. These studies support individual components but do not constitute empirical validation of BF-IJO. The paper proposes judgment-level metrics for false convergence, repeated-reason intervention, evidence-decision alignment, invalid experience rejection, human-gate precision and recall, authority leakage, post-update recurrence, rollback recovery, and audit reconstruction. An EU AI Act appendix provides analytical technical-support scores, not legal conformity assessments.