Kenneth Craik proposed that thinking consists of manipulating internal representations: "If the organism carries a 'small-scale model' of external reality and of its own possible actions within its head, it is able to try out various alternatives, conclude which is the best of them, react to future situations before they arise" (Craik, 1943, p. 61). He identified three processes: (1) translation of external events into internal symbols, (2) derivation of new symbols through inference, and (3) re-translation of symbols back into actions. The fundamental advantage is prediction — "only this internal model of reality...enables us to predict events which have not yet occurred in the physical world." Ding et al. (2025) trace both major contemporary world model lineages back to Craik.

Ding et al. (2025) organize the landscape into two categories: (1) implicit representation — constructing internal models to understand world mechanisms (Ha & Schmidhuber lineage), and (2) future prediction — simulating forward states to guide decision-making (LeCun lineage). Their assessment: "The definition of a world model remains a subject of ongoing debate, generally divided into two primary perspectives." The field exhibits what might be called productive pluralism rather than convergence toward a single definition. Application domains in the survey span games, embodied intelligence, urban intelligence, and societal intelligence — organizational governance is absent.

Ha & Schmidhuber's seminal "World Models" paper defines a world model as learning "a compressed spatial and temporal representation of the environment" enabling an agent to "train entirely inside the 'hallucinated dream' generated by the world model." Their V-M-C architecture (Vision encoder, Memory/predictor, Controller) implements this through a VAE that compresses observations, an MDN-RNN that predicts future latent states, and a remarkably compact linear controller (867 parameters for CarRacing). Critically, they open with Jay Wright Forrester's framing: "The image of the world around us, which we carry in our head, is just a model" — directly citing the founder of System Dynamics and organizational modeling. This bridge between world model research and organizational science has not been developed in subsequent work.

LeCun's position paper proposes Perception, World Model, Cost, Short-Term Memory, Actor, and Configurator modules. The world model's role is "twofold: (1) estimate missing information about the state of the world not provided by perception, (2) predict plausible future states of the world." JEPA (Joint Embedding Predictive Architecture) is "the centerpiece" — predicting in representation space rather than pixel space: "The main advantage of JEPA is that it performs predictions in representation space, eschewing the need to predict every detail of y." LeCun explicitly traces the lineage three times to Kelley-Bryson: "The use of forward models that predict the next state of the world as a function of the current state and the action being considered has been standard procedure in optimal control since the 1950s (Bryson and Ho, 1969)." He positions his architecture "more similar to optimal control than to reinforcement learning." No organizational or governance applications appear anywhere in the paper.

The good regulator theorem proves that any regulator R that is maximally both successful and simple must be homomorphic to the system S being regulated. The proof uses information-theoretic (entropy) arguments: if the regulator achieves minimal outcome entropy H(Z), then the regulator's behavior must be a function of the system's state — H(Z|S) = 0. The simplest such mapping is a homomorphism from S to R. Importantly, the result is a homomorphism, not an isomorphism — the model can be simpler than the system, losing information that is irrelevant to regulation. This means an organizational world model need not capture every detail of organizational reality, only the structural relationships relevant to governance.

The Internal Model Principle proves that for a controller to achieve robust regulation — "closed loop stability and output regulation in the face of small variations in certain system parameters" — it must "incorporate in the feedback path a suitably reduplicated model of the dynamic structure of the disturbance and reference signals." This is both necessary and sufficient. Where Conant-Ashby proves a model is necessary (existence), Francis-Wonham specifies what model is needed (construction): to track a sinusoidal reference, the controller must contain a sinusoidal oscillator; to reject a polynomial disturbance, it must contain integrators of appropriate order. For organizational governance, the implication is that a governance system must contain a model that replicates the dynamic structure of the organizational processes it governs — not just any model, but one with specific structural correspondence.

The Free Energy Principle (Friston et al., 2006; Friston, 2010) states that biological systems minimize variational free energy — "a function of probabilistic beliefs about the world that can be read in terms of self-information." Two mechanisms achieve this: (1) perception/learning (updating beliefs to match sensory evidence) and (2) active inference (acting to make predictions come true). The result is that "the agent becomes a model of the environment in which it is immersed" — the approximate isomorphism between internal model and external contingencies "meets the necessary requirements of Conant and Ashby's Good Regulator Theorem." The FEP is complementary, not competing: it provides a process theory explaining how systems come to satisfy the good regulator theorem, through hierarchical predictive coding and active inference. It extends Conant-Ashby by providing a variational (Bayesian) mechanism for model acquisition, a hierarchical architecture, and a unification of perception, action, and learning.

This three-step chain has not been previously synthesized in the organizational governance literature. Conant-Ashby (necessity) → Francis-Wonham (construction) → Friston (acquisition). Each strengthens the others: Conant-Ashby proves you need a model; Francis-Wonham tells you what it must contain; Friston explains how it gets built and updated. For organizational domains, the chain implies: (1) effective organizational governance is mathematically impossible without a model of organizational dynamics, (2) that model must replicate the dynamic structure of the processes being governed, and (3) the model should be acquired and updated through mechanisms analogous to variational inference — comparing predicted organizational states against actual outcomes and updating accordingly.

Kelley (1960) and Bryson (1962) independently established gradient computation for multi-stage sequential systems. Bryson & Ho (1975) formalized optimal control: given current state and action, compute the optimal next state subject to constraints and cost functions. Sutton (1991) bridged control theory and RL with the Dyna architecture, integrating model learning, planning through simulated experience, and acting. Ha & Schmidhuber (2018) implemented this in deep learning with the V-M-C architecture. LeCun (2022) explicitly acknowledges "The use of forward models that predict the next state of the world as a function of the current state and the action being considered has been standard procedure in optimal control since the 1950s." The independent derivation across distinct fields demonstrates that predict(state, action) → next_state is a mathematical structure, not a paradigm-specific artifact.

Hafner et al.'s DreamerV3 (arXiv 2023, Nature 2025) implements the world model as a Recurrent State-Space Model (RSSM) with encoder, dynamics predictor, reward predictor, continue predictor, and decoder. The latent space uses 32 categorical variables with 32 classes each. The actor-critic trains entirely from imagined trajectories (15-step imagination horizon). Results across 150+ tasks in 7 benchmarks: first algorithm to collect diamonds in Minecraft from scratch (no human data), outperforms specialized methods on Atari, DMLab, and visual control tasks. Critical scaling result: "Increasing the model size of DreamerV3 monotonically improves both its final performance and data-efficiency" — tested across model sizes from 8M to 200M parameters. Symlog predictions and categorical latent spaces eliminate the need for domain-specific hyperparameter tuning.

LeCun's JEPA framework argues for prediction in representation space: "It is important to note that generative latent-variable models are not capable of eliminating irrelevant details, other than by pushing them into a latent variable. This is because they do not produce abstract (and invariant) representations of y." I-JEPA (Assran et al., CVPR 2023) validated this empirically: predicting representations rather than pixels on ImageNet produced highly semantic features without hand-crafted augmentations. V-JEPA (Bardes et al., 2024) extended the principle to video, demonstrating feature prediction as a standalone temporal objective. The architectural principle: what you choose to predict — and at what level — determines what a system can learn. Raw-data prediction wastes capacity on irrelevant detail.

Bengio et al. (2025) define Scientist AI as "a machine that has no built-in situational awareness and no persistent goals that can drive actions or long-term plans" comprising a world model and a probabilistic inference machine. The world model "generates explanatory theories (or arguments, or hypotheses) given a set of observations from the world." A key design choice: "We can compute the probability of a chain of arguments by sequentially multiplying for each argument its conditional probability of being true given the previous arguments are true, which is not possible with the words expressing the arguments. We can thus ensure a clear separation between the probability of an event occurring from the probability of selecting a sequence of words to describe it." The framework distinguishes truth from textual occurrence: "Distinguishing truth from textual occurrence can be done by having the inference machine view X as a latent cause of the observed claim 'someone wrote X,' while also discovering other relevant causes (e.g., the author's intentions)."

The Scientist AI constrains two pillars (redundancy principle): "Although we previously argued that removing a single pillar of agency is sufficient to eliminate agency altogether, we deliberately impose constraints on two. Redundancy is essential in safety protocols, particularly when dealing with a concept like agency, which is not binary but comes in degrees." The Scientist AI has "no situational awareness and no persistent goals" (eliminating goal-directedness) and its actions are "strictly limited to computing probabilistic answers" (constraining affordances). Goal-directedness is defined precisely: "a goal-directed agent is one that breaks an a priori symmetry by preferring one environmental outcome to another (all else being equal)."

Bengio et al. (2023, JMLR) prove that GFlowNets sample compositional objects (sets, graphs) in proportion to a reward function, amortizing MCMC methods into a single trained generative pass. The formal result establishes that maintaining epistemic diversity — multiple competing interpretations — is computationally tractable. The ELK problem (Christiano et al., 2021) identifies a parallel challenge: extracting what a model "actually knows" versus what it reports, when internal representations may contain knowledge that outputs don't surface.

Forrester (1961) founded system dynamics as "an experimental, quantitative philosophy for designing corporate structure and policies." All systems reduce to "two kinds of concepts — levels and flows — and none other." Stocks represent accumulated state; flows embody operating policies assuming "the classic structure of a balancing feedback loop striving to take action to reduce the discrepancy between the observed condition of the system and a goal." The mathematical formalization is differential equations: dS/dt = I(t) - O(t). Sterman (2000) extended this but explicitly rejects prediction-as-forecasting: "All models are wrong." The goal is "understanding structure-behavior relationships rather than relying on forecasts." System dynamics IS explicitly a branch of cybernetics — Forrester drew from Wiener's work and extended Ashby's framework. However, system dynamics models organizational state (stocks) and dynamics (flows) without governance structure: no authority context, no decision lineage, no constraint provenance, no evidence typing. The gap between system dynamics and a formal organizational world model is in governance infrastructure, not in state representation.

A systematic scan reveals: W3C PROV captures data provenance but not organizational state prediction. NIST AI RMF and ISO/IEC 42001 govern AI systems but not organizational world models. 21 CFR Part 11 and SOX govern records and financial controls but not state transformation modeling. PMBOK and PRINCE2 structure project execution without formalizing the prediction function. COSO formalizes internal control but not causal state dynamics. System Dynamics (Forrester, 1961; Sterman, 2000) formalizes organizational state as stocks and flows but lacks governance infrastructure — authority, decision lineage, constraint context. The Balanced Scorecard connects strategy to metrics without temporal state modeling. ERP captures transactional state without decision rationale. The gap is specific: no standard combines state dynamics (which System Dynamics provides) with governance structure (which standards frameworks partially provide) into a unified world model formalization.

Forrester's dictum — "Management is the process of converting information into action" — frames management as a control problem. Budgets predict resource state given planned actions. Forecasts predict future state given current trajectory. Project plans predict completion state given work sequences with resource constraints. Risk assessments predict deviation between expected and actual states. Each performs predict(current_state, action) → future_state without the formal structure that world model research requires. The gap is not that organizations lack prediction practices — it is that their practices lack the causal structure, temporal ordering, authority context, and constraint boundaries that make prediction auditable, queryable, and composable. Ha & Schmidhuber's (2018) citation of Forrester in the opening of their world model paper establishes a direct lineage between organizational modeling and computational world models that neither community has developed.

LeCun maps dual-process theory onto his architecture: "The first mode is similar to Daniel Kahneman's 'System 1', while the second mode is similar to 'System 2'." Mode-1 (reactive policy) operates without world model consultation. Mode-2 (model-predictive control) plans through the world model. Amortized inference bridges them: "This process allows the agent to use the full power of its world model and reasoning capabilities to acquire new skills that are then 'compiled' into a reactive policy module that no longer requires careful planning." The Intrinsic Cost module is explicitly immutable — "To prevent a kind of behavioral collapse or an uncontrolled drift towards bad behaviors, the IC must be immutable and not subject to learning (nor to external modifications)" — functioning as what LeCun calls "safety guardrails." The Configurator, which coordinates all modules, is acknowledged as "the most mysterious" and "least understood."

Hassana Labs (2026), Why World Models Alone Can't Be AGI, proves that world models trained under log loss (or any strictly proper scoring rule) will break physical symmetries that governance-grade systems need to preserve. The core argument: the penalty for maintaining invariance scales with the amount of training data — more data makes the model less likely to preserve structural properties. The compositionality-exchangeability duality shows that compositional structure (needed for causal reasoning) and exchangeability (assumed by standard training) are fundamentally in tension. For organizational domains, the implication is direct: structural properties like authority hierarchies, obligation chains, and constraint boundaries cannot be learned from data — they must be architecturally given. This is the strongest formal argument for why a learned organizational model, no matter how much data it has, cannot substitute for an architecturally specified one.