Three Forms of World Models: Perceptual-Explicit, Relational-Implicit, and Operational-Systemic — Framework and Evidence from an Autonomous Production System

The term "world model" currently denotes three technically distinct classes of objects: perceptual-explicit world models built on explicit 3D representations such as 3D Gaussian Splatting, which are measurable and navigable; relational-implicit world models that emerge inside large language models (LLMs) under next-token prediction pressure and cannot be directly read out; and a third form proposed and validated in this paper—the Operational-Systemic World Model (OSWM): an explicit, queryable, auditable, and hash-anchorable model that an autonomous system maintains of its own component topology, scheduling timeline, state flows, and defect history. We first trace the conceptual lineage of world models through four migrations and establish an orthogonal discriminative coordinate system (modeling target × representation × verifiability). We then report three experiments conducted during 72 hours of continuous auditing on a production autonomous system (973 scripts, 95 scheduled jobs, 33 state namespaces): (1) a peak-shaving simulation driven by the system's model of itself, predicting a peak reduction from 39 to 12 tasks per minute, corroborated by independent probe measurements; (2) a 24-hour invigilated autonomous fault localization and repair trial covering the full pipeline of scanning, target selection, repair, injection assertions, and hash-chain reconciliation; and (3) a dual-source locking experiment on a health metric (independent recomputation 83.94 ≈ 83.9, converting a unilateral claim into dual-source evidence). Counterexample analysis documents two failure modes of the explicit model and how they were caught at the institutional layer. We conclude that for regulated autonomous systems, auditability is a first-order property that outranks accuracy; the trustworthiness of an OSWM derives not from the model itself but from institutions external to it: pre-registered criteria, independent invigilation, and immutable auditing.