Eighty years after Kenneth Craik proposed that the brain carries "small-scale models" of external reality, the term "world model" remains formally undefined. The AI research community uses it to describe everything from variational autoencoders to trillion-parameter architectures dreaming robotic futures, but no authoritative definition specifies what a world model must contain to be complete. The absence has consequences. Without a formal definition, the field cannot distinguish a world model from a prediction engine. It cannot assess completeness. It cannot identify what is missing from a given approach or determine whether two approaches are solving the same problem at different scales.

The result is a landscape of partial implementations, each called a "world model," each covering a different subset of the actual requirements, and none able to explain why its outputs cannot be governed, audited, or trusted in operational contexts. The founding-period research of GrytLabs Research Institute encountered this absence not as a theoretical curiosity but as an engineering constraint: the structural gap identified in TR-A-001 and the architectural necessity argued in TR-A-002 both appeal to a formal definition of what a world model must contain. That definition did not exist in the literature. This report supplies it.

This report makes three contributions. First, a formal definition: ten ingredients that any world model must satisfy, each justified by what becomes impossible without it. Second, a completeness assessment: six leading definitions evaluated against the ten-ingredient standard, establishing that all are incomplete in the same structural pattern. Third, a mechanism taxonomy: three complementary processes — Training, Constitution, and Accretion — by which model ingredients come into existence, establishing that the completeness gap in definitions traces to a mechanism gap in the field's practice.

The thesis would be weakened if: (1) a prior definition were identified that specifies the ten ingredients at comparable formality — meaning the originality claim (WMI-P03) is overstated; (2) an existing definition were shown to be complete against the ten-ingredient standard — meaning the completeness gap does not exist; or (3) training alone were demonstrated to produce constitutive constraints, carried purpose, or attributable memory — meaning the mechanism taxonomy is unnecessary.

The field has been building prediction engines and calling them world models. Prediction is three of ten ingredients.

A world model is the minimum representation that makes a world modelable. It requires ten ingredients: Distinctions (the capacity to differentiate), Entities (things that persist through change), Relations (connective structure between entities), Time (directional ordering), Transition Logic (rules of change, themselves subject to change), Constraints (constitutive limits on what can occur), Observation (structural acknowledgment of partiality), Uncertainty (structural representation of what is unknown), Purpose (the directional bias that makes action evaluable), and Memory (accumulated, accessible history). The compact formula organizes them: State (Distinctions + Entities + Relations) + Change (Transition Logic) + Constraint + Time. Recursion computes the change. Observation, Uncertainty, Purpose, and Memory define the geometry of evolution — from whose perspective the model operates, what it knows it does not know, what it is trying to achieve, and what it has learned from the past.

Six leading definitions are assessed against this standard. Craik (1943) scores 4/10 — the philosophical foundation, but a definition of prediction, not of a world model. Ha & Schmidhuber (2018) score 5/10 — the first computationally realized world model, strong on observation and temporal prediction, but the dream has no rules about what is permissible. LeCun (2022) scores 7/10 — the most complete existing definition, with explicit Memory, Uncertainty, and a Cost Module approaching Purpose, but Constraints remain soft (preferences, not invariants) and Purpose remains designer-injected. DreamZero (2026) scores 5/10 — impressive physical prediction from video, but no governance structure whatsoever. Meta FAIR VLWM (2026) scores 6/10 — notable for acknowledging the constraints gap explicitly ("task-specific penalties or guard-rails" as unspecified inputs) without being able to fill it. The consensus view scores 2/10 — represent, predict, act — Craik's 1943 definition with neural networks attached.

The pattern across all six: Transition Logic (prediction) is universally present. Time and Distinctions are universally present or implied. Constraints, Purpose, and Memory are universally absent or inadequate. The pattern is not random. It maps to a mechanism gap: the field recognizes only one way to build a model — Training — and Training discovers statistical structure from data. Constraints are not statistical regularities. Purpose is not a pattern in data. Memory is not compressed weights. The missing ingredients are missing because the mechanism that could produce them is not recognized.

Three mechanisms are required. Training discovers statistical structure from data — it reliably produces Distinctions, Time, and Transition Logic, partially produces Entities, Relations, and Observation, and cannot produce Constraints, Purpose, Uncertainty (in the epistemic-status sense), or Memory. Constitution specifies structural grammar directly from domain expertise — it produces nine of ten ingredients, including the four Training cannot, but cannot populate Memory with actual operational content. Accretion accumulates governed operational records continuously — it uniquely provides Memory and enriches every other ingredient through operational experience.

The three mechanisms are complementary, not competing. Training provides the intelligence layer. Constitution provides the governance layer. Accretion provides the memory layer. A complete world model requires all three. The field has invested in one. The completeness gap in definitions is a mechanism gap in practice, and the mechanism gap explains why capable systems cannot govern: they are built from one mechanism that cannot produce what governance requires.

This section contextualizes the position statement for proper inspection and interrogation by peers. §3.1 and §3.2 carry the position — the ten-ingredient framework and the three-mechanism taxonomy that constitute the DLP's structural answer to the field's open question. §3.3 carries the conversation — what the field is currently saying, by whom, with what stakes, presented through the completeness assessment of six leading definitions. §3.4 carries the structural question — what is at stake beneath the surface conversation, derived from the field's own findings rather than imposed from outside. The position is stated first because the dialogue is most coherent when the reader can evaluate the field's current state against a named alternative; readers oriented to the conversation may read §3.3 first without loss.

A world model is a representation of a domain that satisfies ten formal requirements. Each ingredient is justified by what becomes impossible without it — not by what it enables (which would make the list aspirational) but by what breaks in its absence (which makes the list structural).

Ingredient 1: Distinctions. The capacity to draw a boundary between this and not-this. Without distinction, there is no information, no structure, no world. Distinction is the founding operation — the act that creates something from nothing by declaring a difference. Every subsequent ingredient presupposes it. Distinctions are not recursive; they are a single operation on which everything else is built.

Ingredient 2: Entities. Things that persist across distinctions — identity maintained through change. An entity is accumulated identity: a person, an organization, an account, a decision, each changing over time while remaining the same thing. Without entities, the model is a field of undifferentiated flux. Entities are recursively defined — an entity can contain other entities at arbitrary depth.

Ingredient 3: Relations. Connective tissue between entities — structure that makes a world a world rather than a heap. Relations are constitutive: an organization is not a set of people but the set of relations between them. Remove the relations and the organization ceases to exist. Relations are recursively structured — a relation can create new relations (a delegation creates an authority chain; a commitment creates an obligation).

Ingredient 4: Time. Ordering with directionality — something came before, something comes after, and that arrow matters. Time provides the axis along which all other recursive processes operate. Without time, the model is a snapshot — it cannot represent change, sequence, cause, or consequence. Time is not recursive; it is the axis for recursion.

Ingredient 5: Transition Logic. The rules by which things change — rules that are themselves subject to change. This is what the field calls "prediction": given a current state and an action, what is the next state? Every existing definition captures this ingredient. But transition logic in a complete world model is recursive: the rules of change are themselves changeable. An organization's decision-making process evolves. A regulatory framework is amended. A model whose transition logic is fixed can predict within a static world; a model whose transition logic is recursive can represent a world that rewrites its own rules.

Ingredient 6: Constraints. What cannot happen. What is conserved. Constitutive limits, not guardrails. Constraints are the most conspicuously absent ingredient in existing definitions. The field treats the world as a space of possibilities and builds models that navigate that space. But the world is equally defined by what is prohibited, and the prohibitions are not preferences or costs — they are constitutive. A chess game is not chess without the constraint that bishops move diagonally; the constraint does not penalize illegal moves but makes them impossible. In organizations, authority boundaries are constitutive constraints (RA-016 §F14, downward causation; RA-016 §F18, five-tradition convergence on bounded nested autonomy). In physics, conservation laws are constitutive constraints. A model that lacks constraints can imagine any future; a model with constraints can distinguish possible futures from permissible ones. This distinction is the foundation of governance. Constraints are recursively tightened and loosened across organizational hierarchies.

Ingredient 7: Observation. A model from somewhere, for someone, with limited information. A world model is not the world — it is always partial, always perspectival, always limited by what the observer can access. This partiality is not a deficiency to be overcome but a structural property that must be represented. A model that cannot represent its own observational limits treats itself as omniscient and contaminates every inference with the assumption of complete information. Observation is recursively refined — each new observation updates the model's representation of its own limits.

Ingredient 8: Uncertainty. What the model does not know, represented structurally. Observation establishes that the model is partial; Uncertainty goes further — the model must carry a structural representation of what is missing. Not just "I haven't seen everything" but "here is what I don't know, here is how confident I am in what I do know, and here is the epistemic status of each claim." In organizational contexts, this is the difference between an audited financial statement, a manager's estimate, and a model's projection — three claims with radically different epistemic status that must be tracked, not collapsed. Uncertainty is recursively reduced through accumulated evidence.

Ingredient 9: Purpose. Teleology — the moment the question "what should I do?" is introduced. Purpose separates a world model from a state description. A state description says "here is how things are." A world model with purpose says "here is how things are, here is how they should be, and here is the distance between the two." That distance drives action, planning, and governance. Every existing definition treats purpose as external — reward signals, cost functions, designer preferences. But in organizations, purpose is constitutive: an organization is its purpose. Remove the purpose and there is no organization (RA-019 §F1, institutional ontology; RA-016 §F16, Morin's dialogical principle). A complete world model carries purpose as a structural ingredient, not as an external input. Purpose is recursively refined through strategic decisions and environmental adaptation.

Ingredient 10: Memory. Retained, accessible history of prior states, transitions, and observations. Without memory, every moment is the first moment. Memory is not storage — it is retained, queryable, accessible history that informs present action. Existing world models have hidden states, context windows, or working memory — none constitute memory in the organizational sense: accumulated, persistent, queryable history that survives beyond the current session (RA-016 §F15, the boundary as first-class subsystem in living systems). Memory is recursively accumulated — every state transition, every decision, every observation adds to it, and the accumulated history shapes future transitions.

A world model requires State + Change + Constraint + Time.

State is composed of Distinctions, Entities, and Relations (Ingredients 1–3). Change is composed of Transition Logic (Ingredient 5). Constraint provides the bounds on change (Ingredient 6). Time provides the axis along which change occurs (Ingredient 4).

Recursion is how the change is computed. Eight of ten ingredients are recursive — they operate on themselves across time, at different rates and different scales.

The remaining four ingredients — Observation, Uncertainty, Purpose, and Memory (Ingredients 7–10) — define the geometry of evolution. They determine from whose perspective the model operates, what it knows it does not know, what it is trying to achieve, and what it has learned from the past. Without them, the model can compute state transitions in a vacuum. With them, the model can govern action in the real world.

The recursion classification is a structural property of each ingredient, not an optional feature:

Eight of ten ingredients are recursive. The two that are not — Distinctions and Time — are the foundational operations on which the recursive structure is built.

In the current AI landscape, "model" and "training" have fused into a single concept. When a researcher, investor, or practitioner hears "model," they understand a parameterized function optimized against a loss function over a dataset. The model exists because it was trained. Training is not merely how the model was built — it is what the model is. The fusion is so complete that it has become invisible, and because it is unstated, its limitations are unexamined.

Training is one of three mechanisms by which models come into existence.

Training exposes a parameterized function to data, optimizes against a loss function, and updates parameters. The model discovers structure from the statistical properties of the data.

Training produces models that capture statistical regularities — distributional patterns, correlations, factual associations, and reasoning patterns. It happens in discrete episodes: train, then deploy. What the model knows is fixed at training time. The structural argument that training cannot preserve declared invariances is established mathematically in RA-018 §F11 (the Hassana Labs symmetry-breaking result: log-loss optimization treats symmetries as degrees of freedom to be exploited for compression, not as structural properties to be preserved).

Training reliably produces three ingredients, partially produces three, and cannot produce four.

Constitution specifies the grammar directly — the primitives, invariants, rules of composition, and structural constraints. The model's architecture is given based on domain expertise, not discovered from data.

Constituted structure is not probabilistic — it does not hold "most of the time." It holds because it was specified to hold. An authority chain exists because someone defined it. A constraint is constitutive because it was declared constitutive.

Constitution produces nine of ten ingredients. It cannot populate Memory — that requires operational experience. The four ingredients that Training cannot produce are precisely the ingredients that Constitution can.

Accretion fills the constituted model through governed operational capture. Every decision, every state transformation, every boundary crossing deposits another layer of organizational truth into the model. The model gets richer not through parameter updates but through the accumulation of governed records.

Accretion happens continuously, at the tempo of organizational operation. The accreted content is the model's living memory — growing monotonically, compounding in value as the accumulated history enables pattern recognition, precedent retrieval, and deviation detection.

Accretion uniquely provides Memory. The three mechanisms together cover all ten ingredients.

A complete world model requires all three. Training provides the intelligence layer. Constitution provides the governance layer. Accretion provides the memory layer. The field has built extraordinary training capabilities and has not built constitution or accretion mechanisms. This is why its models can predict and cannot govern. The deployment-level corollary is empirically documented in RA-020 §F14 (Kellogg et al.'s 80% governance-effort finding: in organizations adopting AI agents, less than 20% of effort is prompt engineering and model development, while over 80% is sociotechnical governance work — the field is paying the mechanism gap in deployment cost).

This subsection presents the conversation the report enters: what the field is currently saying about world models, who is saying it, and where the stakes lie. Six leading definitions of "world model" are evaluated against the ten-ingredient standard. Each is assessed ingredient by ingredient, scored, and analyzed for the structural pattern of its incompleteness. The convergence of incompleteness across six independent definitions is presented as evidence of what the field is currently doing — context for the position stated in §3.1 and §3.2 — not as a convergence-based architectural proof in the TR-A sense.

The philosophical foundation. Craik proposed that the brain carries "small-scale models" of external reality, enabling the organism to "try out various alternatives" and "react to future situations before they arise." The definition covers Time and Transition Logic explicitly, with Distinctions implicit. It lacks Constraints, Uncertainty, Memory, and formalized Purpose entirely.

Completeness: 4/10 (2 explicit, 2 partial, 3 implicit, 3 absent). Craik defines a prediction mechanism, not a world model. Every subsequent definition inherits its gaps because every subsequent definition builds on Craik.

The first computationally realized world model. V (visual encoder) + M (memory RNN) + C (controller). Strong on observation and temporal prediction. The agent can "dream" — train in imagined environments.

Completeness: 5/10 (3 explicit, 3 partial, 4 absent). The dream has no rules about what is permissible. Purpose is externally injected (reward signal), memory decays, relational structure is absent. A prediction engine with a compressed observation model.

The most complete existing definition. A six-module architecture with explicit Perception, World Model, Cost, Short-Term Memory, Actor, and Configurator modules. JEPA as the core architecture — prediction in embedding space.

Completeness: 7/10 (4 explicit, 4 partial, 2 absent). Two critical gaps remain. First, Constraints are soft (costs, preferences) rather than constitutive (invariants, prohibitions). The Cost Module can express that a state is undesirable; it cannot express that a state is impossible. Second, Purpose remains designer-injected; the Cost Module is configured by the system designer, not carried as a property of the domain. In organizational domains where purpose is constitutive, a model that receives purpose externally cannot model an entity whose purpose is intrinsic.

A 14-billion-parameter model that learns physics from video and extracts motor commands from imagined visual futures. Intent specification via language model tokenizer.

Completeness: 5/10 (3 explicit, 2 partial, 5 absent). The clearest illustration of the prediction-engine gap. The system can dream physical futures but cannot specify who authorized the action, under what constraints, or whether the imagined future is permissible. Intent arrives as flat imperative sentences with no governance structure.

A world model predicting goal description, goal interpretation, action description, and world state changes at each planning step. Notable for acknowledging the constraints gap explicitly — "task-specific penalties or guard-rails" referenced as unspecified inputs to the cost function (RA-015 §F4, language-based AI planning analysis).

Completeness: 6/10 (3 explicit, 4 partial, 3 absent). VLWM comes closest to acknowledging that the model requires constraint governance it cannot provide. The four-component prediction structure is structurally convergent with governance architectures designed from practice. But constraints are magical inputs, purpose is task-scoped, and memory is absent.

The informal agreement: a world model should (1) represent the world, (2) predict future states, and (3) enable action based on predictions.

Completeness: 2/10 (1 explicit, 0 partial, 4 implicit, 5 absent). Craik's 1943 definition with neural networks attached. The thinnest possible definition that still uses the term "world model."

The pattern is consistent across all six definitions:

● = explicit ◐ = partial ○ = absent Impl. = implicit

Transition Logic is universally present — every definition includes prediction. Time and Distinctions are universally present or implied. Constraints are absent or partial in every definition. No existing world model includes constitutive limits as a structural ingredient. Purpose is absent or external in every definition. Purpose is injected rather than carried. Memory is absent or ephemeral in every definition. No existing world model accumulates persistent, queryable history.

The universally missing ingredients are the ingredients governance requires. A system without constraints cannot be governed. A system without purpose cannot evaluate whether its actions serve any end. A system without memory operates in an eternal present. The completeness gap in definitions is the governance gap in deployment, and both trace to the mechanism gap in the field's practice.

The mapping from missing ingredients to missing mechanisms is exact, not approximate:

Every missing ingredient is missing because training cannot produce it. Every missing ingredient can be produced by Constitution, Accretion, or both. The field's definitions are incomplete because they assume only one mechanism exists.

This is the structural restatement of Conant and Ashby's (1970) regulator theorem: every good regulator of a system must be a model of that system. A model that cannot constitute its own constraints, carry its own purpose, or accumulate its own memory is not a regulator — it is a predictor of state transitions that has no governance relation to the state being predicted. The training-only paradigm produces predictors; governance requires regulators; the three-mechanism taxonomy names the additional mechanisms (Constitution, Accretion) that turn a predictor into a regulator.

This report establishes the definitional vocabulary — the ten ingredients and three mechanisms — for the World Model Initiative thesis. Its claims are structural and definitional, not empirical. The ten-ingredient framework is proposed as a formal standard against which any world model definition can be evaluated. The completeness assessment is conducted against that standard. The mechanism taxonomy is proposed as a structural explanation for the completeness gap.

Irreducibility. The report claims ten irreducible ingredients but does not formally prove irreducibility — that no ingredient can be derived from the combination of others. A reduction test (attempting to derive each ingredient from the remaining nine) is identified as an open question. The justification offered is structural: each ingredient is justified by what becomes impossible in its absence, and the absences are distinct. A model without Constraints faces a different failure mode than a model without Purpose, which faces a different failure mode than a model without Memory. Distinct failure modes support irreducibility but do not prove it.

Exhaustiveness. The report does not formally prove that ten ingredients are sufficient — that no additional ingredient is required. Domain-specific applications may surface requirements the current framework does not name. The claim is that these ten are necessary, not that they are sufficient. The open question of whether additional ingredients exist is identified for future work.

Mechanism boundaries. The three-mechanism taxonomy draws clear boundaries (Training discovers, Constitution declares, Accretion accumulates) that real systems may blur. Fine-tuning is a training episode that can incorporate constituted knowledge. Retrieval-augmented generation approximates memory without formal accretion. The taxonomy describes structural categories, not implementation boundaries.

Assessment methodology. The completeness assessment evaluates published definitions and architectures. Implementation details not described in the published work may satisfy ingredients the assessment marks as absent. The assessment scores what is documented, not what may exist in undisclosed implementations.

Mathematical formalism. The compact formula (State + Change + Constraint + Time) is presented as a structural organization of the ten ingredients, not as a mathematical formalism. Formal mathematical grounding is deferred to RA-022 (compiler theory grounding).

The framework would be weakened or falsified by:

1. Reduction. A demonstration that any ingredient can be derived from the combination of others, reducing the count below ten.
2. Completeness counterexample. An existing definition or system that satisfies all ten ingredients, contradicting the claim that no existing system is complete.
3. Mechanism counterexample. A demonstration that training alone can produce constitutive constraints, carried purpose, or attributable memory — that the mechanism taxonomy is unnecessary.
4. Prior art. A prior definition specifying the same ingredients at comparable formality, contradicting the originality claim (WMI-P03).

The list is aspirational, not structural. The ten ingredients could be read as a wish list of desirable properties rather than irreducible requirements. The counter-evidence is the justification-by-absence: each ingredient is justified by what breaks without it, not by what it enables. The failure modes are distinct and operationally documented across the research record.

The mechanism gap is temporary. Training may eventually produce all ten ingredients as scale, data, and architecture improve. The counter-evidence is the mechanism analysis: log-loss optimization is structurally incapable of distinguishing rarity from prohibition, confidence from provenance, or statistical memory from queryable history. The limitation is mathematical, not empirical — it does not attenuate with scale.

The assessment is uncharitable. The completeness scores undercount what existing definitions actually achieve. The counter-evidence is that each assessment scores what the published definition specifies, with partial credit for ingredients that are acknowledged but not formalized. The assessment errs on the side of inclusion — LeCun's Cost Module receives partial credit for Constraints despite encoding preferences rather than invariants.

The evidence reviewed in this report supports the following position statement dispositions under the World Model Initiative thesis.