The Theoretical Starting Point and Progressive Research Agenda of MRAC: From Non-Convergence Diagnosis to Engineering AGI

Abstract

Previous work introduced Multi-Round Audit Convergence (MRAC), which evaluates whether a system can stably handle a current task by tracking how issues change across repeated “audit–remediation–re-audit” cycles on the same engineering object. Another study explained uncertainty in AI delivery through the relationship among Search Space (S), Coverage Ability (C), and Delivery Reliability (R). Taking these two studies as premises, this paper examines what kinds of coverage judgments can be supported by MRAC non-convergence and convergence, respectively, and organizes them into a progressive L1–L5 research agenda. Based on explicit operational definitions and applicability conditions, this paper argues that, when task and audit conditions remain stable, identified issues are valid, remediations are continuously subjected to re-audit, and resource boundaries are specified in advance, persistent MRAC non-convergence can serve as a conditional diagnosis that the current complete AI development system cannot yet reliably cover the task search space. For L2, when remediation results can be preserved and drift is controlled, convergence indicates that exposed failure paths have been more thoroughly handled and validated, and provides positive evidence that local coverage under the current audit configuration is approaching saturation. The representativeness of this evidence for the complete delivery-relevant search space, and the strength of any extrapolation from it, still require empirical validation. On this basis, L3 studies whether recursive decomposition can restore convergence for non-convergent tasks, L4 examines whether complete engineering workflows can advance autonomously over extended periods, and L5 tests whether similar mechanisms can extend to intelligent tasks beyond engineering.