The Inquiry
The Inquiry: Does the state-prediction formalism established across world model research apply to organizational domains, and if so, does a formal organizational world model standard exist?
Falsifiable formulation: 1. The state-prediction formalism `predict(state, action) → next_state` is domain-general — derivable independently from optimal control theory, model-based RL, cognitive architecture, and cybernetics. 2. No formal world model standard exists for organizational reality — distinct from data provenance (W3C PROV), AI governance (NIST AI RMF), and audit infrastructure (SOX, 21 CFR Part 11). 3. The Conant-Ashby good regulator theorem and the Francis-Wonham internal model principle together establish that effective organizational governance requires a model of organizational dynamics — not as metaphor but as mathematical necessity. 4. Existing organizational modeling traditions (System Dynamics) formalize organizational state as stocks and flows but do not provide governance-grade structure (authority, decision lineage, constraint context). 5. The convergent critique from three independent research directions (LeCun, Bengio, Hassana Labs) identifies structural requirements any formal organizational world model must satisfy.
Executive Summary
### The Forrester Bridge — A Connection Neither Community Has Developed
Ha & Schmidhuber (2018) open their seminal world model paper with Forrester's (1971) observation that "The image of the world around us, which we carry in our head, is just a model." This is not incidental framing — Forrester founded System Dynamics as the application of cybernetics and control theory to organizational management. Yet neither the world model research community nor the System Dynamics community has developed this bridge. World model research proceeded into games, robotics, and autonomous driving. System Dynamics proceeded into policy simulation and organizational design. The shared intellectual ancestor (Wiener's cybernetics, Ashby's variety, and the state-prediction formalism) sits at the root of both traditions, but no one has formally connected them for organizational governance. Ding et al.'s (2025) survey covers games, embodied intelligence, urban intelligence, and societal intelligence — organizational governance is absent. The gap is not a gap in individual papers but a gap in the field's collective attention.
### The Good Regulator Derivation Chain — From Necessity to Construction
The Conant-Ashby → Francis-Wonham → Friston chain is the strongest theoretical argument for why organizational world models are necessary, not merely useful. Conant-Ashby proves any good regulator must be a model (necessity). Francis-Wonham proves the model must replicate the dynamic structure of the governed processes (construction). Friston proves that systems acquire such models through prediction-error minimization (acquisition). Applied to organizations: effective governance requires a model of organizational dynamics; that model must replicate the causal structure of organizational state transformations; and the model improves through deviation measurement — comparing predicted states against actual outcomes.
This chain is more powerful than the prediction-formalism argument (F9) alone because it proves necessity, not just applicability. The prediction formalism shows that the world model concept can be applied to organizations. The good regulator chain shows that effective organizational governance must involve a model — with or without a formal protocol implementing it. Organizations that govern effectively are, by the theorem, already modeling their dynamics — just informally.
The homomorphism result (Conant-Ashby) is specifically significant: the organizational world model does not need to capture every detail. It needs structural correspondence at the level relevant to governance. This is the organizational equivalent of LeCun's JEPA insight — predict at the right abstraction level, not at raw data level.
### System Dynamics as the Near-Miss — What It Gets Right and What It Lacks
System Dynamics is the closest existing formalization to an organizational world model. It formalizes organizational state (stocks), dynamics (flows), feedback structure (reinforcing and balancing loops), and the connection between information and action (Forrester: "Management is the process of converting information into action"). It is explicitly cybernetic. It uses differential equations to model state evolution. It has 60+ years of application.
What System Dynamics lacks for a formal organizational world model: (1) Authority context — System Dynamics models what happens and why (feedback structure) but not who decided or under what authority. (2) Decision lineage — state transitions in System Dynamics are governed by rate equations, not by traceable decision chains with evidence and constraint context. (3) Governance structure — System Dynamics models the system being governed but not the governance process itself. (4) Epistemic typing — Sterman acknowledges "all models are wrong" but the framework has no mechanism for typing the epistemic status of model elements. (5) Multi-model coordination — System Dynamics assumes a single modeler's view; it has no mechanism for multiple competing models under authority-governed arbitration.
These five gaps define the specific contribution space for a formal organizational world model. The contribution is not state dynamics (System Dynamics does this) but governance infrastructure over state dynamics.
### The Event/Description Separation — Bengio's Key Insight for Organizational Governance
Bengio's (2025) separation of "the probability of an event occurring" from "the probability of selecting a sequence of words to describe it" has an underappreciated parallel in organizational governance. Organizations routinely confuse descriptions of state with state itself: budgets (descriptions of intended resource allocation) are treated as resource allocation; strategic plans (descriptions of intended direction) are treated as direction; risk registers (descriptions of perceived threats) are treated as threat assessment. The gap between the description and the described organizational state is exactly where governance fails — and it is exactly the gap that Bengio identifies as requiring formal architectural treatment.
This connects to the ELK problem (Christiano et al., 2021) in organizational form: a populated organizational record contains latent knowledge about organizational dynamics that no individual has articulated because no individual holds the full picture. The descriptions (budgets, plans, reports) are what individuals produce. The organizational state — the actual configuration of authority, resources, obligations, and constraints — is the latent variable that descriptions approximate but do not capture.
### The Convergent Critique as Structural Requirements
Three independent research programs converge on necessary conditions that any adequate world model must satisfy:
1. Abstraction level (LeCun): Prediction must operate at a representation level that eliminates irrelevant detail — not at raw data level. 2. Causal structure (Bengio): The model must express cause-and-effect relationships as logical statements with computable probabilities — not just associative patterns. 3. Imposed invariances (Hassana Labs): Structural properties that governance requires must be architecturally given — training objectives under standard loss functions will break them.
For organizational domains, all three conditions apply simultaneously, and they are harder to satisfy than in physical domains because organizational state is linguistically constituted (declarations create obligations, authorizations create authority, agreements create commitments) rather than physically observable.
Findings26
F-RA-008-01 · theoretical-grounding · established
Craik (1943) originated the world model concept — internal "small-scale models" of external reality enabling prediction — an insight that has remained stable for eight decades; he identified three processes (translation, derivation, re-translation) with prediction as the fundamental advantage. Ding et al. (2025) trace both major contemporary world-model lineages back to Craik.
F-RA-008-02 · gap-identification · lab-originated
The field agrees on the function of world models (state prediction) but not on what they are; Ding et al.'s 2025 ACM survey divides the landscape into implicit-representation (Ha & Schmidhuber lineage) and future-prediction (LeCun lineage) and characterizes this as "productive pluralism," with organizational governance absent from surveyed application domains.
F-RA-008-03 · structural-mapping · lab-originated
Ha & Schmidhuber (2018) established the modern computational paradigm (V-M-C architecture: VAE compression, MDN-RNN prediction, compact linear controller) and opened the paper by citing Jay Wright Forrester — founder of System Dynamics / organizational modeling — establishing a bridge to organizational science that subsequent work has not developed.
F-RA-008-04 · theoretical-grounding · established
LeCun (2022) proposed a six-module cognitive architecture (Perception, World Model, Cost, Short-Term Memory, Actor, Configurator) centered on JEPA (prediction in representation space), explicitly tracing the lineage three times to the Kelley-Bryson optimal-control tradition and positioning the architecture "more similar to optimal control than to reinforcement learning"; no organizational/governance applications appear.
F-RA-008-05 · formal-establishment · lab-originated
Conant & Ashby (1970) proved that any regulator R that is maximally both successful and simple must be homomorphic to the system S it regulates (good regulator theorem, information-theoretic proof: H(Z|S)=0 → behavior is a function of system state); the result is a homomorphism, not an isomorphism — the model may be simpler than the system.
F-RA-008-06 · design-requirement-derivation · lab-originated
Francis & Wonham (1976) provided the constructive engineering formalization — the Internal Model Principle: for robust regulation a controller must "incorporate in the feedback path a suitably reduplicated model of the dynamic structure of the disturbance and reference signals" (necessary and sufficient). To track a sinusoid the controller must contain an oscillator; to reject a polynomial disturbance, integrators of appropriate order.
F-RA-008-07 · root-cause-diagnosis · lab-originated
Friston's Free Energy Principle (Friston et al., 2006; Friston, 2010) provides the complementary process theory: biological systems minimize variational free energy via perception/learning and active inference, such that "the agent becomes a model of the environment" — meeting the requirements of Conant-Ashby's theorem; it adds a variational (Bayesian) acquisition mechanism, a hierarchical architecture, and unification of perception/action/learning.
F-RA-008-08 · formal-establishment · lab-originated
The Conant-Ashby → Francis-Wonham → Friston chain establishes a complete mathematical argument — governance requires models (necessity), the models must have specific structural properties (construction), and variational inference supplies the acquisition mechanism — a synthesis not previously made in the organizational governance literature.
F-RA-008-09 · theoretical-grounding · established
The state-prediction formalism `predict(state, action) → next_state` was independently derived across optimal control (Kelley 1960, Bryson 1962, Bryson & Ho 1975), model-based RL (Sutton 1991, Dyna), and deep learning (Ha & Schmidhuber 2018), with LeCun (2022) acknowledging it as standard in optimal control since the 1950s.
F-RA-008-10 · empirical-demonstration · established
DreamerV3 (Hafner et al., 2023/2025) validates the predict-imagine-act paradigm at scale — an RSSM-based world model with categorical latent space; actor-critic trained entirely from imagined trajectories; first to collect diamonds in Minecraft from scratch; across 150+ tasks in 7 benchmarks, increasing model size (8M→200M params) monotonically improves both final performance and data efficiency.
F-RA-008-11 · design-requirement-derivation · lab-originated
Prediction at the right abstraction level determines efficiency and semantic quality: LeCun's JEPA argues for prediction in representation space; I-JEPA (Assran et al., CVPR 2023) validated semantic feature learning from representation (not pixel) prediction on ImageNet; V-JEPA (Bardes et al., 2024) extended this to video. Raw-data prediction wastes capacity on irrelevant detail.
F-RA-008-12 · design-requirement-derivation · lab-originated
Bengio's Scientist AI (Bengio et al., 2025) separates "the probability of an event occurring" from "the probability of selecting a sequence of words to describe it," computing probabilities over chains of logical arguments rather than word sequences, and treating X as a latent cause of the observed claim "someone wrote X."
F-RA-008-13 · theoretical-grounding · established
Bengio defines agency through three pillars — Affordances, Goal-Directedness, Intelligence — and argues eliminating any one is sufficient to eliminate danger; the Scientist AI constrains two (redundancy principle): no situational awareness / no persistent goals (eliminating goal-directedness) and actions "strictly limited to computing probabilistic answers" (constraining affordances).
F-RA-008-14 · gap-identification · lab-originated
GFlowNets (Bengio et al., 2023, JMLR) formalize diverse hypothesis generation — sampling compositional objects in proportion to a reward function, amortizing MCMC into one trained generative pass — proving epistemic diversity is computationally tractable; the ELK problem (Christiano et al., 2021) identifies the parallel challenge of extracting what a model actually knows versus what it reports.
F-RA-008-15 · gap-identification · lab-originated
System Dynamics (Forrester, 1961; Sterman, 2000) formalizes organizational state as stocks and flows with feedback (dS/dt = I(t) − O(t)) and is explicitly cybernetic, but explicitly rejects prediction-as-forecasting ("all models are wrong") and lacks governance structure — no authority context, decision lineage, constraint provenance, or evidence typing. The gap is in governance infrastructure, not state representation.
F-RA-008-16 · gap-identification · lab-originated
No formal organizational world model standard exists that combines state representation with governance structure: W3C PROV (provenance, not state prediction), NIST AI RMF / ISO 42001 (govern AI, not org world models), 21 CFR Part 11 / SOX (records/financial controls), PMBOK / PRINCE2 (project execution), COSO (internal control, not causal state dynamics), System Dynamics (state, no governance), Balanced Scorecard (strategy-metrics, no temporal state), ERP (transactional state, no rationale).
F-RA-008-17 · gap-identification · lab-originated
Every organizational prediction practice (budgets, forecasts, project plans, risk assessments) is an informal, unstructured instance of `predict(current_state, action) → future_state` — and Forrester identified management as "the process of converting information into action" in 1961; these practices lack the causal structure, temporal ordering, authority context, and constraint boundaries that make prediction auditable, queryable, and composable.
F-RA-008-18 · theoretical-grounding · established
LeCun's System 1 / System 2 framework with amortized inference provides a formal model for progressive automation: Mode-1 reactive policy (no world-model consultation) vs Mode-2 model-predictive planning; amortized inference "compiles" deliberative skills into reactive policy; the Intrinsic Cost module is explicitly immutable ("safety guardrails"); the Configurator (coordinating all modules) is "the most mysterious" / "least understood."
F-RA-008-19 · record-correction · lab-originated
Yang's cost-of-action argument is a workshop contribution (Mila, Feb 2026, founder-attended), NOT a published finding — Yang et al. (2023) defines sequential decision making via standard MDP formalism without critiquing it, and its listed open problems do not include action cost or MDP limitations; the cost-of-action insight ("a world model of not just the outcome but also the time it takes") comes exclusively from her workshop panel remarks.
F-RA-008-20 · convergent-validation · lab-originated
Four independent research programs converge on the insufficiency of current world-model approaches, each addressing a different structural layer: LeCun (architecture) — predict in representation space; Bengio (structure) — impose explicit causal/compositional priors; Yang (formalism) — MDP omits action cost/time; Chlon (objective) — training under log loss breaks symmetries, so invariances must be architecturally imposed. The four critiques are non-redundant and compositionally exhaustive (where to compute / what structure / what to represent / why optimization alone fails).
F-RA-008-31 · architectural-resolution-claim · lab-originated
Hassana Labs (2026), *Why World Models Alone Can't Be AGI*, proves that world models trained under log loss (any strictly proper scoring rule) break physical symmetries governance-grade systems must preserve — the invariance penalty scales with training-data volume (more data → less likely to preserve structure), and a compositionality-exchangeability duality shows compositional structure (causal reasoning) and exchangeability (standard training) are fundamentally in tension. Therefore governance invariances (authority hierarchies, obligation chains, constraint boundaries) must be architecturally imposed, not learned.
F-RA-008-32 · structural-mapping · lab-originated
[Synthesis] Ha & Schmidhuber's citation of Forrester constitutes an undeveloped bridge between world-model research and organizational science: both traditions descend from a shared cybernetic ancestor (Wiener's cybernetics, Ashby's variety, the state-prediction formalism), yet neither the AI/ML world-model community nor the System Dynamics community has connected them — and the absence is a gap in the field's collective attention, not in individual papers (confirmed by Ding et al.'s and Yang et al.'s surveys, where organizational governance is missing).
F-RA-008-33 · formal-establishment · lab-originated
[Synthesis] The good-regulator chain proves organizational world models are *necessary*, not merely *applicable* — a strictly stronger claim than the prediction-formalism argument (F9) alone: organizations that govern effectively are, by the theorem, already modeling their dynamics (informally), with or without a formal protocol.
F-RA-008-34 · root-cause-diagnosis · lab-originated
[Synthesis] Bengio's event/description separation has an underappreciated organizational parallel: organizations routinely confuse descriptions of state (budgets, strategic plans, risk registers) with state itself, and this description-vs-described gap is exactly where governance fails; the ELK problem recurs in organizational form — a populated organizational record holds latent knowledge no individual has articulated because no individual holds the full picture.
F-RA-008-35 · convergent-validation · lab-originated
[Synthesis] The convergent critique defines necessary structural conditions any adequate world model must satisfy — abstraction level (LeCun), causal structure (Bengio), action-cost representation (Yang), imposed invariances (Chlon) — and for organizational domains all conditions apply simultaneously and are *harder* to satisfy than in physical domains because organizational state is linguistically constituted (declarations create obligations, authorizations create authority, agreements create commitments) rather than physically observable.
F-RA-008-36 · gap-identification · lab-originated
[Conclusion] Progressive governance automation maps onto amortized inference, but the Configurator (LeCun's "most mysterious," "least understood" module) maps to the organizational meta-governance question — "what governs the governance?" — which remains open.
Bibliography21
Assran, Mahmoud and Duval, Quentin and Misra, Ishan and Bojanowski, Piotr and Vincent, Pascal and Rabbat, Michael and LeCun, Yann and Ballas, Nicolas (2023) · Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture
Bardes, Adrien and Garrido, Quentin and Ponce, Jean and Chen, Xinlei and Rabbat, Michael G. and LeCun, Yann and Assran, Mahmoud and Ballas, Nicolas (2024) · Revisiting Feature Prediction for Learning Visual Representations from Video
Bengio, Yoshua and others (2025) · Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path?
Bengio, Yoshua and Lahlou, Salem and Deleu, Tristan and Hu, Edward J. and Tiwari, Mo and Bengio, Emmanuel (2023) · GFlowNet Foundations
Bryson, Arthur E. and Denham, Walter (1962) · A Steepest-Ascent Method for Solving Optimum Programming Problems
Bryson, Arthur E. and Ho, Yu-Chi (1975) · Applied Optimal Control: Optimization, Estimation and Control
{Hassana Labs} (2026) · Why World Models Alone Can't Be {AGI}
Christiano, Paul and Cotra, Ajeya and Xu, Mark (2021) · Eliciting Latent Knowledge: How to tell if your eyes deceive you
Roger C. Conant and W. Ross Ashby (1970) · Every Good Regulator of a System Must Be a Model of That System
Craik, Kenneth J. W. (1943) · The Nature of Explanation
Ding, Kangrui and others (2025) · Understanding World or Predicting Future? A Comprehensive Survey of World Models
Forrester, Jay W. (1961) · Industrial Dynamics
+9 more citations