Movendi Research Note No. 2

From World Models to Reality Models

Abstract

Movendi is initiating a foundational research program to define and develop Reality Models: computational systems that learn and maintain a compact, persistent, and predictive internal representation of a bounded physical environment.

The program begins from the hypothesis that constraining an environment spatially, operationally, temporally, and dimensionally may materially change the nature of the machine-learning problem. A Reality Model does not initially attempt to represent the entire physical world. It seeks to discover and model the variables that determine the evolution of a particular environment or class of environments.

This research will investigate whether physical intelligence can be achieved through systems that are smaller, more specialized, more computationally efficient, and more continuously adaptive than current general-purpose foundation models. It will also question whether several dominant foundations of contemporary machine learning, including tokenization, global back-propagation, transformer attention, predominantly static model parameters, and the separation of training from inference; are necessarily the most appropriate foundations for persistent physical intelligence.

This note presents the scientific context of the program, proposes a provisional definition of a Reality Model, and introduces Movendi's primary research hypothesis. It does not prescribe a particular architecture. Existing world-model and machine-learning architectures will be treated as experimental baselines rather than predetermined solutions.

Foreword

Movendi Research Note No. 1 introduced our foundational hypothesis that intelligence emerges from prediction and described the Direct Predictive Model as a possible approach to learning the dynamics of physical reality directly from experience.

This second note narrows and extends that hypothesis.

The question is no longer only whether a machine can learn predictive representations of reality. It is also what form of reality should first be modeled, what information the model must retain, what it should attempt to predict, and whether the dominant foundations of contemporary machine learning are appropriate for this task.

Our initial answer is to constrain the problem deliberately.

Rather than beginning with an open-ended representation of the world, we propose to study bounded physical environments whose relevant dimensions can be characterized, observed, and progressively learned. We call the resulting systems Reality Models.

This definition remains provisional. It is an exercise intended to establish a sufficiently precise scientific object around which hypotheses can be formulated, experiments designed, and results evaluated.

As with all Movendi Research Notes, this document is being published before the research has produced final conclusions. Its purpose is to expose the reasoning behind the program, invite criticism, and support collaboration with researchers, students, and institutions interested in predictive physical intelligence.

Executive Summary

Current world-model research broadly seeks to develop systems that construct internal representations of an environment and predict how that environment may evolve.

Many of these efforts pursue increasingly general models capable of representing open-ended environments. This is an important direction, but it creates a learning problem of extraordinary dimensionality.

The observable physical world contains more information than any model can completely represent. Much of the information present in sensory data has little relevance to a particular prediction. Different future states may appear visually plausible while being physically inconsistent. Uncertainty may result from incomplete observation, hidden variables, unknown actions, stochastic processes, or limitations in the model itself.

Generative quality alone does not resolve these problems. A model may generate a convincing image of the future without maintaining a coherent representation of the physical state that produced it.

Movendi is approaching the problem from a different direction.

A Reality Model is not initially intended to represent the entire physical world. It models a dimensionally constrained physical environment: a bounded domain in which the relevant objects, actions, processes, temporal horizons, observations, and physical or operational constraints can be characterized.

Examples could include a tabletop workspace, a laboratory station, an industrial cell, a hospital room, a vehicle interior, or a robotic workspace.

The physical size of the environment is only one possible constraint. A room may contain an enormous number of possible states and interactions, while a much larger industrial process may be highly constrained by its equipment, operating modes, permitted actions, and known physical relationships.

The principal hypothesis is that reducing the effective dimensionality and ambiguity of the environment may make robust predictive modeling possible with substantially less data, compute, and model capacity than would be required for an open-world model.

For the purpose of this research program, we provisionally define a Reality Model as:

A computational system that learns and maintains a compact, persistent, and predictive internal representation of a constrained physical environment; estimates its present state; predicts its possible future evolution; and accounts for the effects of actions, interactions, incomplete observation, and uncertainty.

This definition is intentionally architecture-agnostic.

A Reality Model may ultimately use transformers, learned state-space models, recurrent architectures, object-centric representations, predictive coding, geometric models, dynamical systems, explicit memory, deterministic transition functions, or combinations of these methods.

The research program will not begin by assuming that tokenization, global attention, end-to-end backpropagation, static model structures, or monolithic neural architectures are necessary.

Existing approaches will provide essential baselines. The objective is to determine experimentally which foundations remain appropriate for physical intelligence, which require modification, and which may need to be replaced.

The most ambitious possible outcome would not simply be a better world model. It would be evidence for a distinct class of artificial intelligence designed specifically to learn, maintain, and predict bounded physical reality.

Scientific Context

The challenge of open-ended world models

A system that perceives the present but cannot predict the future has only a limited representation of the environment in which it operates.

Prediction requires more than identifying visible objects or generating likely future images. It requires an internal representation that preserves the variables governing how a physical state can change.

In an open-ended environment, the possible state space is extremely large. A sensory stream may contain information about geometry, color, texture, illumination, motion, sound, materials, objects, people, equipment, weather, and countless other variables.

Only a subset of these variables may be relevant to a particular physical transition.

A model predicting whether an object will fall may need to represent its support, center of mass, geometry, motion, and external forces. It may not need to reconstruct every visible texture or lighting variation in the scene.

The difficulty is determining which information is relevant before the model has fully learned the dynamics of the environment.

Open-ended models must also contend with several distinct sources of uncertainty:

• Observations may be incomplete or noisy.
• Relevant variables may be hidden.
• An action may not yet have been selected.
• External interventions may be unknown.
• A process may contain genuinely stochastic behavior.
• The model may lack an appropriate representation.
• The model may simply not have observed enough examples.

These are materially different conditions. A model that combines them into a single distribution may produce plausible predictions without understanding why those predictions remain uncertain.

Increasing generality also has a computational cost. A broadly capable model must allocate capacity to a vast number of objects, contexts, actions, and relationships that may never be relevant within a particular physical environment.

Movendi's research program does not assume that this generality is always necessary.

Constraining the problem

A Reality Model begins with a bounded environment.

The constraint may include:

• Physical space and geometry.
• Types of objects or agents.
• Available sensing modalities.
• Possible actions and interactions.
• Relevant temporal horizons.
• Physical or logical processes.
• Operational states and objectives.
• Expected categories of change.
• Required predictions and outputs.

This does not mean that every future event must be specified in advance. Nor does it mean that the environment must be simple.

It means that the model operates within a defined domain whose effective dimensionality may be discovered and progressively represented.

A constrained environment could still contain humans, novel events, incomplete observations, and complex physical interactions. The distinction is that the model does not need to preserve every possible dimension of the broader world. It needs to preserve the dimensions that affect the evolution of its environment.

Some of those dimensions may be specified through instrumentation or prior knowledge. Others may need to emerge from prediction.

The model may discover:

• Which observations contain predictive information.
• Which sensory details are redundant.
• Which objects, surfaces, or relationships persist.
• Which actions materially alter future state.
• Which temporal scales govern different processes.
• Which hidden variables explain apparently uncertain transitions.
• Which aspects of the environment can be represented deterministically.
• Which information can be discarded without reducing predictive performance.

The objective is not merely to compress sensory data.

It is to identify the smallest—or at least a sufficiently compact—internal state that preserves the information required to predict the future dimensions of the environment that matter.

This leads to the first foundational proposition of the program:

Constraint may not merely make the same modeling problem smaller. It may create a materially different problem for which a more specific and efficient form of artificial intelligence becomes possible.

An Exercise in Definition

The term world model has been used to describe a broad range of systems, from internal representations used in reinforcement learning to generative models capable of predicting or synthesizing future observations.

Movendi uses the term Reality Model to identify a narrower scientific object.

A Reality Model is not defined only by its input modality, architecture, or ability to generate a visual future. It is defined by its relationship to a specific physical environment and by its ability to maintain and predict the state of that environment over time.

Our provisional definition is:

A computational system that learns and maintains a compact, persistent, and predictive internal representation of a constrained physical environment; estimates its present state; predicts its possible future evolution; and accounts for the effects of actions, interactions, incomplete observation, and uncertainty.

Each part of this definition introduces a research requirement.

A computational system

The definition does not prescribe a neural network, transformer, or any other specific architecture.

The final system could combine learned representations, explicit memory, geometric structures, analytical models, deterministic functions, probabilistic components, or computational mechanisms not yet identified.

The purpose of the definition is to characterize what the system must accomplish, not how it must be implemented.

It learns

A Reality Model should not depend entirely on a manually engineered description of its environment.

Existing physical and operational knowledge may be incorporated where useful, but the model should be capable of discovering representations and dynamics from observation, interaction, and prediction error.

The long-term objective is a system that progressively improves its representation of the environment with limited explicit supervision.

It maintains

Physical reality persists between observations.

An object does not necessarily cease to exist when it moves outside a camera's field of view. A machine does not lose its state because a telemetry packet is delayed. A person temporarily hidden behind another surface remains part of the environment.

A Reality Model must therefore maintain state rather than merely process independent observations.

It should be able to retain prior information, revise earlier estimates, preserve identity, and distinguish between what is directly observed and what remains inferred.

Its representation is compact

A useful internal state should not contain every detail of every sensory observation.

The representation should retain the variables required for prediction while discarding information that has no material effect on the future states of interest.

Compactness is therefore not only a question of compression ratio. It concerns state sufficiency: whether the representation contains enough information to predict future evolution without reconstructing the entire history of sensory inputs.

Its representation is persistent

The internal state should survive across time, interruptions, partial observations, and temporary occlusions.

Persistence may require several forms of memory:

• Current environmental state.
• Historical or episodic memory.
• Object-specific memory.
• Stable environmental knowledge.
• Temporary hypotheses.
• Unresolved uncertainty.

A foundational research question will be which information should reside in model parameters and which should remain in an explicit, updateable memory.

Its representation is predictive

The internal state should be evaluated according to what it enables the system to predict.

A representation that reconstructs a visually accurate observation but fails to preserve the variables governing future evolution may not constitute an effective Reality Model.

Prediction may involve:

• The next environmental state.
• Changes to an existing state.
• Object or agent trajectories.
• Events and their timing.
• Consequences of actions.
• Violations of physical or operational constraints.
• Multiple possible future states.
• The model's confidence in those futures.

The appropriate prediction target remains a central research question.

It represents a constrained physical environment

A Reality Model is attached to a defined physical domain.

Its internal state should correspond, directly or indirectly, to persistent elements of that environment rather than only to statistical regularities in an undifferentiated observation stream.

This attachment to bounded physical reality differentiates the proposed research object from a model trained primarily to generate plausible sensory outputs across open-ended domains.

It estimates the present state

Observation and state are not equivalent.

A camera image, point cloud, microphone signal, or telemetry value is a measurement of reality, not reality itself. Measurements may be partial, noisy, delayed, or contradictory.

The model must infer a present state from the available evidence and retain distinctions between:

• What has been observed.
• What is currently inferred.
• What is remembered.
• What remains unknown.
• What has been predicted.

It predicts possible future evolution

A Reality Model should represent not only what exists, but how that state may change.

Some transitions may be highly predictable. Others may depend on hidden variables, human choices, or future actions. The model may therefore need to produce one expected state, several conditional states, or a structured range of possible futures.

The objective is not necessarily perfect long-horizon prediction. It is an increasingly accurate and calibrated representation of how uncertainty develops over time.

It accounts for actions and interactions

Physical environments do not evolve only through passive dynamics.

Humans, robots, machines, software systems, and environmental processes intervene. A useful model may need to distinguish autonomous evolution from action-conditioned change.

Whether actions should be represented as inputs, inferred as hidden causes, or modeled through separate components remains an open research question.

It represents uncertainty

A Reality Model should not hide uncertainty behind a single plausible output.

It should attempt to distinguish between:

• Measurement uncertainty.
• Hidden-state uncertainty.
• Unselected or unknown actions.
• Multiple physically plausible futures.
• Stochastic processes.
• Gaps in the model's own knowledge.

This distinction may become essential for dependable prediction and future control.

What a Reality Model Is Not

A Reality Model may incorporate perception, digital twins, generative models, simulation, and control systems. It is not defined by any one of them.

A perception system may identify what is currently visible without maintaining a persistent predictive state.

A digital twin may represent an environment through manually engineered geometry, states, and rules without learning those representations directly from experience.

A generative model may produce visually convincing future observations without preserving object identity, physical consistency, or causal structure.

A control system may determine what action should be taken without maintaining a sufficiently complete model of the physical environment.

A Reality Model may eventually interact with all of these components. Its defining function is to learn and maintain the predictive state on which they can operate.

Visualization is a readout, not the objective

A Reality Model should be capable of providing a visual representation of its estimated present state and, where useful, its predicted futures.

This is important for inspection, interpretation, human interaction, debugging, and scientific evaluation.

However, photorealistic rendering should not be the primary measure of success.

A visually convincing prediction may conceal an inadequate physical representation. Conversely, an abstract latent, geometric, or object-centric representation may accurately preserve the variables that govern future evolution.

Visualization should therefore be treated as a readout of the model's internal state, not as the state itself.

Primary Research Hypothesis

The proposed research program is based on the following hypothesis:

Dimensionally constrained physical environments may permit a distinct class of artificial intelligence whose representations, learning mechanisms, memory systems, and computational architecture are better adapted to persistent physical prediction than those of current general-purpose foundation models.

Such a system may be:

• Smaller.
• More specialized.
• More sample-efficient.
• More computationally efficient.
• Persistent and stateful.
• Continuously adaptive.
• More predictable within its domain.
• Better able to separate deterministic structure from uncertainty.

This hypothesis does not presume that contemporary machine-learning foundations are unsuitable.

It proposes that they should be tested rather than treated as axioms.

The research program will therefore examine whether physical intelligence necessarily requires:

• Tokenized representations.
• Transformer-based global attention.
• End-to-end global backpropagation.
• Large and predominantly static parameter sets.
• Knowledge stored primarily in model weights.
• A strict separation between training and inference.
• A single monolithic architecture.
• Probabilistic treatment of all state transitions.
• Increasing scale as the principal path to capability.

The strongest result may be an appropriately adapted transformer or state-space architecture.

It may be a hybrid system combining learned models with persistent memory, geometry, analytical dynamics, deterministic transitions, or explicit causal structure.

It may also be a substantially different computational architecture.

The program should remain capable of producing — and recognizing, each of these outcomes.

Existing Foundations as Baselines, Not Conclusions

Tokenization

Tokenization has enabled machines to process language and other modalities as sequences of discrete representational units.

Physical environments, however, evolve continuously in space and time. Their persistent structures may be better described through objects, surfaces, fields, trajectories, events, relationships, or dynamical states.

The research will examine whether tokenization remains an effective computational abstraction, whether it must be modified, or whether it imposes artificial boundaries on continuous physical processes.

Backpropagation

End-to-end backpropagation has been central to the progress of contemporary machine learning.

A continuously operating Reality Model may nevertheless require several forms of learning at different speeds.

Correcting the estimated position of an object should not necessarily require retraining a global network. Learning a new property of one environment may differ from improving the shared foundations of the Core Engine.

The program will investigate whether adaptation can be divided among:

• Immediate state correction.
• Explicit memory updates.
• Local component learning.
• Environment-specific adaptation.
• Periodic global consolidation.
• Changes to general model knowledge.

Backpropagation will remain an important baseline, but global weight updates will not be assumed to be the only valid learning mechanism.

Transformers and attention

Transformers provide powerful methods for discovering relationships across sequences and modalities.

Physical interactions are often local, sparse, persistent, and structured. A model may not need to compare every element of an environment with every other element during every update.

Recurrent, hierarchical, spatially local, event-driven, modular, object-centric, or dynamical architectures may offer different advantages.

The research question is not whether transformers work. It is whether they provide the most appropriate and efficient foundation for persistent physical state.

Memory and model parameters

General knowledge, environmental state, and episodic history are different forms of information.

Learning that a particular object has moved should not necessarily require changing the foundational model. It may require updating an explicit memory associated with that object and environment.

The program will investigate how information should be divided among:

• General model parameters.
• Environment-specific knowledge.
• Current state.
• Episodic memory.
• Object-specific memory.
• Stable physical relationships.
• Temporary hypotheses.

Training and inference

Conventional machine learning generally separates training from deployment and inference.

A Reality Model may instead operate through a continuous cycle:

1. Observe.
2. Estimate the current state.
3. Predict.
4. Compare the prediction with subsequent reality.
5. Attribute the error.
6. Correct the state.
7. Adapt memory or the model where justified.

The important question is not whether every observation should change the model. It is which part of the system should change, from what evidence, at what rate, and under which safeguards.

The Research Program

Movendi is initiating a multi-year research program to investigate these questions with academic researchers and doctoral candidates.

The initial program will seek to:

• Formalize the Reality Model problem.
• Define bounded experimental environments and evaluation methods.
• Establish strong contemporary architectural baselines.
• Investigate representations for persistent physical state.
• Develop self-supervised and continuously adaptive learning methods.
• Test alternatives or complements to tokenization, transformer attention, and global backpropagation.
• Design the foundations of a reusable Reality Model Core Engine.
• Train and evaluate a first basic Reality Model.
• Investigate deterministic, conditionally deterministic, and uncertain dynamics.
• Determine which components can transfer between constrained environments.

The program is intended as foundational research rather than outsourced product development.

Its objective is not to force a predetermined architecture toward a commercial demonstration. It is to establish a scientific basis for understanding what a Reality Model must represent, how it should learn, and which computational foundations are best suited to the problem.

Future Research Notes and Experiment Reports will describe the research framework, experimental environments, baselines, hypotheses, evaluation methods, and progressive development of the Reality Model Core Engine.

Why Now?

The recent development of large-scale self-supervised and generative models has demonstrated that machines can learn rich representations from enormous quantities of data.

At the same time, the computational requirements and architectural complexity of increasingly general models continue to grow.

Movendi's experience deploying physical AI systems in manufacturing and healthcare has approached the problem from the opposite direction. These systems combine perception, digital twins, telemetry, state machines, machine-learning models, control systems, and human-defined knowledge to maintain operational representations of physical environments.

They can solve valuable problems, but their representations are substantially engineered. Every new environment requires integration, modeling, configuration, and domain-specific logic.

This leads to a more fundamental question:

Can a machine learn the predictive state of a constrained physical environment directly from observation and interaction, rather than relying on an increasingly complex assembly of manually integrated representations?

The tools required to investigate this question are becoming available.

The scientific opportunity is to determine whether the next step should be another increase in scale — or a deliberate reduction in scope.

Toward a More Specific Form of Intelligence

The prevailing direction in artificial intelligence has been toward increasingly general models trained on increasingly broad datasets.

Reality Models explore a complementary direction.

How much intelligence can emerge from a system that is deliberately specific?

How much smaller can a model become when its environment is bounded?

Can specialization produce greater reliability with less compute?

Can a model discover the dimensions that govern future evolution and discard those that do not?

Can deterministic physical structure be learned rather than approximated through an undifferentiated distribution of plausible futures?

Can a system improve continuously without repeatedly retraining a large, static model?

The intended result is not simply a smaller general-purpose model.

It is a computational system designed around the requirements of bounded physical reality: persistent state, incomplete observation, physical causality, continuous adaptation, and prediction over time.

Reality Models begin by accepting a limit. They do not initially attempt to understand everything.

Our hypothesis is that this limit may be precisely what allows them to understand a physical environment deeply enough to predict it.