Formalizing Autonomous Task Admission in Multimodal Agents: A Reliability-Based Meta-Controller for Latent World Models

This preprint introduces cECT (computable Embodied Constructor Theory) as a reliability-based meta-controller for multimodal agents and latent world models. The central problem addressed is the Capability Admission Gap: Large Multimodal Models (LMMs) can generate plausible plans, tool-use strategies, and action descriptions, but they lack an internal criterion for determining when such generated behavior corresponds to an actually executable and reliable capability. The paper formalizes a System-2 control layer over the System-1 generative substrate of LMMs. This layer is defined by an operational quadruple: Q_t, M_t, ω_t, and KP_t. Q_t functions as a latent world model or state predictor; M_t functions as an inertial action-bias state; ω_t functions as a dynamic regulatory gain; and KP_t functions as a Knowledge Potential, interpreted here as an evidence accumulator for reliable system-environment coupling. The key engineering contribution is the use of KP_t as a task-admission gatekeeper. Rather than allowing an agent to treat a fluent plan, a single successful rollout, or a plausible tool-use sequence as evidence of real competence, the cECT controller admits a candidate task τ into the agent’s repertoire P_t only when preconditions, reliability, and KP_t thresholds are satisfied. This reframes grounding in multimodal agents as validated transformation rather than representation matching. The paper reports an initial sufficiency experiment in a hidden-structure environment. In this configuration, the observable state was capped below 0.46 before unlock, while the target task required stabilization around 0.75 ± 0.03. The cECT engine detected a latent gate, crossed the KP threshold at step 21, admitted the out-of-distribution task at step 22, and reproduced successful task admission across 20/20 validation reruns. The result is interpreted as sample-efficient autonomous task admission in fewer than 25 interactions under the tested configuration. The broader aim of the paper is to provide an auditable control architecture for agentic AI systems that distinguishes representation from reliable transformation. LMMs may generate candidate plans and actions, but a reliability-based System-2 controller is required to validate which transformations have actually become executable capabilities.