The World Model Definition — Ten Ingredients, Three Mechanisms, and a Completeness Assessment of Leading Approaches

The term "world model" has been in use for eighty years without a formal definition specifying what a world model must contain to be complete. This report proposes one. A world model is the minimum representation that makes a world modelable — not a simulation, not a control system, not a dynamic system, but the more primitive thing those depend on. Ten ingredients are identified as irreducible requirements: Distinctions, Entities, Relations, Time, Transition Logic, Constraints, Observation, Uncertainty, Purpose, and Memory. Six leading definitions — Craik (1943), Ha & Schmidhuber (2018), LeCun (2022), DreamZero (2026), Meta FAIR VLWM (2026), and the informal consensus — are evaluated against these requirements. All are incomplete in precisely the same ways: the universally missing ingredients are Constraints, Purpose, and Memory, with Uncertainty and Relations absent or partial in most. The pattern is not random. The missing ingredients map exactly to the missing mechanism: the field recognizes only one way to build a model — training — and training cannot produce what it cannot find in data. A three-mechanism taxonomy (Training, Constitution, Accretion) identifies the complementary processes required to satisfy all ten ingredients. The completeness gap in existing definitions is a mechanism gap in the field's practice. The definition and taxonomy together establish the formal vocabulary for the World Model Initiative thesis.