Predictive Artificial Intelligence Architectures — Executive Summary

Initial Observation

Large Language Models (LLMs) represent a major breakthrough in artificial intelligence. They show that large-scale learning from language can generate coherent text, assist programming, answer questions, summarize documents and orchestrate external tools.

However, this achievement should not be confused with complete general intelligence. Language is a trace of the world; it is not the world itself. Human and animal intelligence learn through continuous experience involving perception, action, memory, error correction, social interaction, causality and abstraction.

LLMs can learn useful internal representations, including spatial and temporal structures. Nevertheless, these representations often remain fragile, format-dependent and insufficient for embodied, robust and planning-oriented understanding. See Gurnee & Tegmark, Berglund et al. and Bender et al..

Proposed Analytical Framework

This memorandum now develops a broader central axis: Predictive Artificial Intelligence Architectures. It does not treat World Models as an exclusive doctrine, but rather as one major family of solutions within a wider architectural framework.

The objective is to analyze how an AI system can remember, abstract, predict, reason causally, plan, act and remain governable.

Predictive representations may take several forms, including explicit World Models, causal models, experiential memories, symbolic planners, tool-using agents, active inference systems, neuro-symbolic architectures and cyber-physical control loops.

The decisive debate is therefore not simply “World Model or not?”. It is rather the following question: which predictive architecture, at which level of abstraction, with which memory, which causality, which capacity for action and which security control?

The Role of World Models

The term “World Model” remains an important reference. It belongs to a scientific tradition rooted in the mental models of Craik, causal models in cognitive science, model-based reinforcement learning described by Sutton & Barto, the World Models of Ha & Schmidhuber, and later JEPA / V-JEPA architectures associated with LeCun, Bardes et al. and Assran et al..

In this memorandum, the World Model becomes one pillar among others, rather than the sole interpretive center.

The general conclusion is that the most credible path will probably be hybrid: language, perception, memory, causality, symbolic reasoning, external tools, predictive models, planning, action, cybersecurity, identity and trust governance.

Freemindtronic Positioning

The trajectory involving the Cryptographic Genome, EviDNA and Digital DNA through CryptPeer/EviSKMS industrialization is documented in a distinct companion memorandum.

This approach assumes an inventor-researcher posture grounded in applied observation, continuous monitoring of the state of the art, identification of weak and strong signals, analysis of hardware and software attack vectors, and the design of sovereign counter-espionage, encryption, authentication and trust solutions.

This field experience does not replace scientific evaluation. It provides the empirical starting point for a vision that must be formalized, protected, compared and tested.

The DNA/EviDNA companion memorandum documents the observable industrialization of EviSKMS-CryptPeer through verifiable elements, including trusted runtime, Runtime Integrity, DRT continuity, RSCC, fail-closed policies, anti-replay mechanisms, chained logs, cryptographic governance, sovereign passwordless operation, DDNA Gen1, security testing campaigns and deployment artifacts.

This appendix discloses no source code, pseudo-code, internal formats or transition rules, in order to preserve current and future intellectual property protections.

The industrial trajectory also relies on an internationally patented foundation: the Segmented-Key Authentication System (FR3063365 B1, WO/2018/154258 family and EP, US, CN, JP and KR extensions).

This granted title enables limited public discussion of the published principles of cryptographic segmentation, physical proximity and conditional trust reconstitution, without exposing Gen2 genomic extensions, the complete DRT engine or EviSKMS mechanisms developed after the founding patent.

Patent / industrialization / confidential tripartition (Register A). Patent WO/2018/154258 constitutes a public document of prior art and technological foundation. CryptPeer/EviSKMS industrialization is documented through observable industrialization elements and non-sensitive evidence (Register A). Genomic extensions, internal mechanisms and unpublished know-how remain protected under Register C.

A chain of time-stamped public disclosures from 2018 to 2026 is listed in the companion memorandum.

For the public reference publication, the present memorandum includes a section on limitations, falsifiability and scope of validity, as well as a short version.

Cryptographic details and CNRS/EviDNA comparisons are addressed in the companion memorandum. These additions aim to distinguish what is demonstrated, what is industrialized, what belongs to applied research and what remains open to independent validation.

Predictive Artificial Intelligence Architectures — Foundational Thesis of the Memorandum

This memorandum proposes the following formulation as its scientific foundation.

This thesis deliberately shifts the focus of the memorandum. The central issue is no longer to defend one specific school of thought or to oppose LLMs to World Models. The subject becomes broader: identifying the functions required for a robust predictive intelligence architecture.

These functions are: understanding language, perceiving or integrating context, remembering experience, abstracting relevant variables, anticipating possible evolutions, reasoning causally, planning, acting and remaining controllable.

An LLM may be excellent at language. A symbolic engine brings particular strength to formal logic. A causal model clarifies intervention and counterfactual reasoning. A World Model helps predict the evolution of an environment, while active inference seeks to reduce uncertainty through action.

None of these approaches, taken in isolation, is sufficient today to constitute a robust general intelligence.

The research question therefore becomes: how can these capabilities be composed into a coherent, verifiable, secure architecture capable of long-term learning?

Methodological Note — Inventor-Researcher Posture and Applied Observation

This reflection is not limited to a conventional academic approach. It is also grounded in a long inventor-researcher experience, built over more than fifteen years through continuous observation of digital threats, analysis of weak and strong signals, study of hardware and software attack vectors, and the design of digital counter-espionage, encryption, authentication and sovereign trust solutions.

This applied observation posture has gradually led to a core conviction: digital security can no longer be reduced to isolated mechanisms of protection, identification or compliance. It must be understood as a continuity of trust capable of linking identity, context, proof, memory, governance, hardware environment, software runtime and the evolution of threats over time.

The memorandum therefore assumes a dual nature. It engages with the scientific state of the art while also carrying a vision derived from invention, industrialization and operational analysis of attack surfaces. This articulation between documentary research, field observation and technical design forms the basis of the Freemindtronic trajectory around EviSKMS, CryptPeer as an industrialized embodiment of this approach, and the Cryptographic Genome as a conceptual and prospective formalization.

This posture does not claim to replace scientific validation with individual experience. It clarifies the origin of the approach: an architectural hypothesis born from prolonged observation of threats, reinforced by the industrialization of solutions, then formulated as a research framework intended to be compared, evaluated and discussed.

Eurosatory 2022–2026 Trajectory — From Human DNA to the Cryptographic Genome

This inventor-researcher posture was built through successive public milestones. Presentations delivered at Eurosatory between 2022 and 2026 help clarify the evolution of the reasoning, from an initial cyber foundation toward a trust architecture based on human DNA, then toward the Cryptographic Genome as a response to the time factor.

In 2024, this trajectory reached an industrial milestone with DataShielder Defense NFC HSM. The product does not only address the highly secure sharing of cryptographic keys associated with DNA. It also introduces an initial continuity of operational identity. The person who creates the key knows to whom it is transmitted, the recipient holds a trusted NFC device, and importing the key into that device establishes a controlled relationship between identity, physical possession, cryptographic key and encrypted/signed use.

This continuity, however, remains tied to a hardware and transactional perimeter involving the NFC device, the compatible terminal, key validity, media governance and control over the sharing lifecycle. It therefore provides a partial response to trust over time, without fully covering the challenge of a durable, re-evaluable and governable identity in a context where current AI, and later predictive AI architectures, may transform recognition, authentication, decision and trust models.

This shift led, in 2026, to the demonstration of Digital DNA and the Cryptographic Genome Generator. Biological DNA remains usable, but it becomes one possible element within a broader structure designed to organize proof continuity, trust criteria, segmentation, traceability, governance and the evolution of identity over time. This transition is therefore not a rupture. DataShielder Defense NFC HSM provides operational identity continuity, while Digital DNA and the Cryptographic Genome extend this approach toward a durable, contextualized, re-evaluable and governable identity. This evolution constitutes one of the cyber-identity application cases of the present memorandum on Predictive Artificial Intelligence Architectures.

Strict Definitions

To avoid ambiguity, this memorandum uses the following definitions.

Artificial General Intelligence. The ability of a system to learn, reason, plan and act across diverse domains, including novel situations, with robustness and adaptability beyond simple pattern memorization.

World Model. An explicit or implicit internal representation that enables a system to predict the evolution of an environment, especially under the effect of possible actions. See Craik, Ha & Schmidhuber and the World Model for Robot Learning Survey.

Predictive Representation. An internal structure that is not only used to recognize a situation, but also to anticipate its future transformations.

Causality. The ability to distinguish a correlation from a productive mechanism and to reason about what would happen under intervention. See Pearl and Schölkopf et al..

Planning. The ability to evaluate several possible sequences of actions, simulate their consequences and select a trajectory aligned with a goal.

Experiential Memory. A form of memory that stores not only documents or facts, but also episodes, errors, strategies, abstractions and feedback that can be reused for future action. See Du.

Grounding. The relationship between symbols, language, perception, action and environment. The symbol grounding problem is discussed by Harnad.

Predictive Artificial Intelligence Architectures — Introduction

Contemporary artificial intelligence is advancing at an unprecedented pace, driven largely by Large Language Models (LLMs). These systems generate text, answer questions, summarize documents, translate languages, write code, and assist users across a wide range of intellectual tasks.

Their impressive performance can sometimes create the impression that they are approaching artificial general intelligence. This perception, however, deserves careful examination. Large Language Models are trained primarily on vast collections of textual data. They learn to predict the most probable continuation of a sequence by identifying statistical regularities within their training data.

Although this capability is remarkable, it does not necessarily imply a deep understanding of the world. Language describes objects, events, intentions, relationships, and causes, yet it cannot replace perception, action, sensory feedback, or embodied experience.

From the earliest stages of life, humans learn through vision, movement, touch, interaction with objects, the consequences of their actions, social relationships, and continuous experience of the physical world. They progressively construct abstractions, discard irrelevant details, and retain the structures that enable prediction and purposeful action.

The central question therefore becomes: can robust intelligence emerge solely from learning based on text? Or does it require a new generation of architectures capable of integrating language, memory, abstraction, causality, prediction, action, and Trust Governance?

This dissertation adopts the following position: Large Language Models are indispensable, yet they are likely insufficient on their own. World Models represent an important direction, but not the only one. The next stage of AI should instead be conceived as a convergence of Predictive Artificial Intelligence Architectures combining World Models, Neuro-symbolic AI, Causal Models, Active Inference, Tool-using Agents, Agentic Memory, AI Planning, AI Cybersecurity, and Cyber-Physical Trust.

1. Large Language Models — Capabilities and Limitations

Large Language Models are trained on enormous quantities of textual data. Their training corpora may encompass a substantial portion of publicly available Internet content, supplemented by additional sources such as books, scientific articles, web pages, document repositories, software code, and annotated conversations.

The model transforms this massive body of information into internal parameters. This process can be viewed as a form of statistical compression of human language. Rather than memorizing every sentence, the model learns structures, associations, stylistic patterns, grammatical regularities, factual knowledge, and recurring reasoning patterns present within its training data.

This approach enables remarkable performance. LLMs can explain concepts, solve certain classes of problems, reformulate ideas, generate coherent text, and orchestrate external tools. Nevertheless, their operation remains fundamentally grounded in predicting the most probable continuation of a sequence of text.

This limitation explains several well-known challenges, including hallucinations, the absence of native persistent memory, the fragility of certain forms of generalization, difficulty with long-horizon planning, and the lack of direct grounding in the physical world. The critique proposed by Bender et al. reminds us that language alone does not guarantee situated understanding.

A balanced scientific position is therefore not to claim that LLMs never reason. Rather, it is more precise to state that LLMs can produce useful reasoning and acquire certain forms of Predictive Representations of the world, yet these representations are not currently sufficiently stable, causal, embodied, or verifiable to constitute a complete artificial general intelligence.

Predictive Artificial Intelligence Architectures — 2. What LLMs Already Do Well

A rigorous scientific dissertation should not caricature Large Language Models. They are far more than statistical dictionaries. They can learn abstract regularities, perform reasoning expressed in natural language, generate software code, manipulate mathematical representations, invoke external tools, and occasionally infer information that is not explicitly stated in a prompt.

Research such as Gurnee & Tegmark suggests that some language models internally encode spatial and temporal dimensions as exploitable latent structures. This finding indicates that learning from text alone can give rise to latent Predictive Representations of the world.

However, these internal representations should not be confused with robust World Models. The Reversal Curse, for example, demonstrates that a model may learn a relationship in one direction while failing to generalize correctly to the inverse relationship. This fragility suggests that certain capabilities remain strongly dependent on the training distribution and on how a problem is formulated.

The scientific question is therefore not, “Do LLMs understand or not?” Rather, it is: What internal representations do they construct, under which conditions, how robust are these representations, and to what extent can they support causality, Agentic Memory, AI Planning, and purposeful action?

3. The Real Cost of Modern Artificial Intelligence

Massive investment in artificial intelligence is driven primarily by two requirements: computational infrastructure and post-training.

The first concerns the computing resources required to train and deploy modern AI models. Training relies on specialized processors, memory, networking, energy, and large-scale data centers. Inference at scale is also computationally expensive, as every request consumes resources while imposing constraints on latency, availability, and security.

The second concerns post-training. A raw model is not automatically reliable, useful, or safe. It must be refined through supervised learning, human feedback, alignment, filtering, instruction tuning, tool integration, retrieval mechanisms, and security policies.

This reality demonstrates that the raw model alone is insufficient. Modern artificial intelligence already depends on an ecosystem composed of models, data, external memory, tools, guardrails, interfaces, governance policies, infrastructure, and continuous supervision.

This observation reinforces the central thesis of this dissertation: advanced AI will likely not consist of a single isolated model, but rather of a composite architecture.

Predictive Artificial Intelligence Architectures — 4. Human Learning: Sensory Experience, Action, and Abstraction

Comparing a Large Language Model with a young child highlights the fundamental difference between text-based learning and embodied learning.

By the age of four, a child has already experienced thousands of hours of wakefulness. During this period, the child continuously receives visual, auditory, tactile, and motor sensory inputs. The retina does not transmit a raw image directly to the brain; it transforms, filters, and compresses information before sending it through the optic nerve. Although estimates vary, the literature on retinal coding indicates that the transmitted information flow remains substantial. See Koch et al..

Any comparison with the token streams processed by Large Language Models must therefore remain cautious. Human visual experience should not be presented as directly equivalent to textual data. The essential qualitative distinction is that a child learns from a continuous, active, multimodal sensory experience that is intrinsically linked to the consequences of its own actions.

The child observes objects, interacts with them, experiences the consequences, adjusts expectations, memorizes regularities, and progressively constructs abstractions. The child learns that some objects fall, roll, break, resist, disappear behind others, or reappear. At the same time, the child acquires an understanding of intentions, social cues, emotions, and implicit rules.

Human intelligence is therefore not built simply through the accumulation of information. It emerges through experience, interaction, abstraction, prediction, and error correction. This perspective aligns with the work of Lake et al., who emphasize the importance of Causal Models, intuitive physics, intuitive psychology, and rapid generalization.

5. Why Text Alone Is Likely Not Enough

Text is a secondary representation of the world. It describes objects, events, emotions, intentions, and relationships. It is not the world itself.

A model trained exclusively on text learns the regularities of language about the world, but not necessarily the regularities of the world itself. It may learn that people often write “a glass falls and breaks,” yet this knowledge remains mediated through language. It does not arise from direct experience of gravity, fragility, sound, trajectory, or the physical consequences of an action.

This distinction is closely related to the Symbol Grounding Problem discussed by Harnad. A symbol cannot be considered fully understood if it is connected only to other symbols. At least part of its meaning must ultimately be grounded in perception, action, or experience.

This does not imply that text is of limited value. On the contrary, language is an extraordinarily powerful medium for abstraction, cultural transmission, and reasoning. However, language alone appears insufficient to produce robust embodied intelligence.

The most scientifically accurate formulation is therefore that text alone can give rise to rich internal Predictive Representations, yet it does not appear sufficient to build a general intelligence capable of perception, causality, Experiential Memory, AI Planning, and effective action within the physical world.

Predictive Artificial Intelligence Architectures — 6. World Models as a Family of Predictive Architectures: Origins of the Concept

Throughout this dissertation, World Models are no longer treated as the exclusive focus of the discussion. Instead, they are examined as one of the major families of Predictive Artificial Intelligence Architectures because they explicitly formalize a fundamental capability: anticipating how an environment evolves from its current state under possible actions.

The term World Model is not a recent invention. It extends a long-established scientific tradition.

Craik proposed that the mind constructs small-scale internal models of reality, enabling actions to be mentally simulated before being executed. This intuition remains fundamental: thinking consists, in part, of testing possible actions internally before acting in the external world.

Johnson-Laird later developed the theory of mental models, according to which human reasoning relies on internal representations of possible situations.

In artificial intelligence, the concept emerged through model-based reinforcement learning, in which an agent uses a model of environmental dynamics to simulate the consequences of alternative actions. See Sutton & Barto.

The expression World Models became explicit in the work of Ha & Schmidhuber, who learned compressed representations of environments and used them to train autonomous agents. More recently, the JEPA and V-JEPA architectures proposed by LeCun, Bardes et al., and Assran et al. have extended this approach by learning to predict abstract latent representations rather than individual pixels.

The concept itself is therefore not new. What is new is its renewed central role in contemporary discussions about the future of artificial intelligence.

7. World Models as Predictive Representations: A Rigorous Definition

A World Model is a specific form of Predictive Representation: an internal representation that enables a system to predict how an environment is likely to evolve.

Within the broader perspective of this dissertation, it is not presented as the only solution, but rather as a central class of architecture capable of linking state, action, future, and decision-making.

Formally, if a system observes the state of the world at time t, denoted x_t, it constructs an abstract representation s_t. Given a candidate action a_t, the model predicts a future state s_{t+1} or a probability distribution over possible future states.

Observation x_t
      ↓
Encoder E
      ↓
Abstract state s_t
      ↓ + action a_t
Predictor P
      ↓
Predicted future state ŝ_{t+1}

The value of a World Model lies not merely in recognizing the current state of the environment, but in predicting what may happen next.

A system equipped with a World Model can answer the question: What would happen if I performed this action? This question lies at the heart of AI Planning, practical causality, and autonomous intelligence.

Predictive Artificial Intelligence Architectures — 8. Abstraction and Hierarchical Representations

It is impossible to represent the complete state of the world down to its ultimate physical details. Describing an ordinary room at the level of quantum field theory would be impractical: the complete wave function of a macroscopic system cannot be measured, and no realistic computation could predict every physically possible evolution in a useful manner.

Humans do not reason in this way. Instead, they construct abstractions: objects, surfaces, agents, intentions, obstacles, trajectories, rules, tools, and risks. Each level of abstraction discards part of the lower-level detail while preserving the information that is useful for prediction at a given scale.

This hierarchy mirrors the organization of science itself: particle physics, nuclear physics, chemistry, biochemistry, molecular biology, biology, psychology, sociology, and ecology. Each discipline focuses on the level of organization most relevant to its domain.

An effective World Model must therefore learn hierarchical representations. Lower levels may encode shapes, textures, and motion. Intermediate levels may encode objects, relationships, and scenes. Higher levels may encode intentions, constraints, goals, norms, and abstract causal structures.

Intelligence does not consist of preserving every detail, but of constructing the appropriate level of abstraction for effective action.

9. Learning Through Prediction: Encoder, Predictor, and Prediction Error

A system can learn a World Model through self-supervised predictive learning.

1. It observes the world at time t as input data x_t.
2. An encoder transforms x_t into an abstract representation s_t.
3. A predictor estimates the future state ŝ_{t+1}.
4. The system then observes x_{t+1}.
5. The same encoder produces the corresponding representation s_{t+1}.
6. The system minimizes the discrepancy between ŝ_{t+1} and s_{t+1}.

The objective is to make predictions within a meaningful abstract representation space rather than predicting every individual pixel. This is precisely the intuition behind JEPA architectures: learning to predict useful representations instead of reconstructing every detail. See LeCun and Bardes et al..

This learning mechanism fundamentally changes the nature of learning: the system no longer learns merely to recognize the world; it learns to anticipate its evolution.

Predictive Artificial Intelligence Architectures — 10. From Prediction to Planning

Planning requires the ability to simulate multiple possible futures.

To choose an action, an intelligent agent must evaluate several possible trajectories:

Current state
   ├── Action A → Possible future A
   ├── Action B → Possible future B
   └── Action C → Possible future C

The agent then compares these alternative futures according to an objective, a constraint, a cost, or a risk.

This capability is fundamental to model-based reinforcement learning, where an internal model is used to simulate the consequences of actions before they are executed. See Sutton & Barto.

Planning may also be delegated to symbolic engines, constraint solvers, search trees, or formal verification tools. Even in these cases, however, the system must represent states, actions, and transitions. In other words, planning almost always reintroduces some form of World Model.

11. World Models Within Predictive Artificial Intelligence Architectures: Promise and Limitations

This section retains World Models as a major scientific reference while placing them within a broader architectural perspective. Their importance lies not in belonging to a particular school of thought, but in the function they embody: learning useful Predictive Representations that support prediction, AI Planning, and action.

11.1. Generative World Models

The World Models proposed by Ha & Schmidhuber learn a compressed representation of an environment and then use this representation to train an autonomous agent. This approach demonstrates that an agent can learn not only from the real or simulated world, but also from an internally learned model of that world.

Predictive Artificial Intelligence Architectures — 11.2. JEPA, V-JEPA, and Latent-Space Prediction

JEPA and V-JEPA architectures aim to predict abstract representations rather than individual pixels. Their objective is to capture the information that is most relevant for understanding and action, without expending learning capacity on secondary visual details. See LeCun, Bardes et al., and Assran et al..

11.3. World Models in Robotics

World Models have become a major research direction in robotics because they enable systems to predict environmental dynamics, simulate actions, perform planning, and improve out-of-distribution generalization. See World Model for Robot Learning Survey.

11.4. Embodied Robotics and Digital Simulators

Simulation environments and digital twins make it possible to generate rare or hazardous scenarios. They are particularly valuable for autonomous driving, industrial robotics, and physical agents. Nevertheless, a simulation is never a complete representation of the real world and must always be validated against real-world observations.

Predictive Artificial Intelligence Architectures — 11.5. Limitations of World Models

World Models are not a universal solution. They face several fundamental challenges:

• learning stable abstractions;
• managing uncertainty and multiple possible futures;
• distinguishing Causal Models from statistical correlations;
• avoiding the prediction of irrelevant details;
• generalizing beyond the training distribution;
• integrating language, action, and Agentic Memory;
• evaluating model quality objectively;
• ensuring safety when actions affect the physical world.

An inaccurate World Model may become dangerous precisely because it appears internally coherent. Consequently, evaluation, Trust Governance, and Runtime Integrity become central requirements.

Predictive Artificial Intelligence Architectures — 12. Competing and Complementary Approaches

This section maps the principal research directions pursuing the same overarching objective: robust reasoning, generalization, AI Planning, memory, hallucination reduction, and reliable action.

12.1. Neuro-symbolic AI

Neuro-symbolic AI combines neural networks with symbolic reasoning, including rules, logic, knowledge graphs, constraint solvers, and inference engines.

This approach is particularly promising in domains where explainability, verification, and regulatory compliance are essential, including law, AI Cybersecurity, mathematics, formal verification, diagnostics, Trust Governance, and safety-critical systems. See Garcez & Lamb, Colelough & Regli, and Yang et al..

Primary strength: explainable and controllable reasoning. Main limitation: limited grounding in perception and the physical world. Relationship to World Models: symbolic systems often perform planning over abstract states and therefore frequently reintroduce a discrete or logical World Model.

12.2. Tool-using LLM Agents, RAG, Memory, and Planning Systems

A major industrial direction consists of using Large Language Models as orchestration engines. They invoke external tools, retrieve information, execute code, consult knowledge bases, rely on external Agentic Memory, and delegate specialized tasks to dedicated modules.

Retrieval-Augmented Generation (RAG) improves factual accuracy by connecting the model to external knowledge sources. See Lewis et al.. Tool-using Agents further extend LLM capabilities through planning, reasoning, tool use, and memory. See Yao et al., Huang et al., and Du.

Primary strength: immediate operational effectiveness. Main limitation: retrieval and external tools do not replace genuine causal understanding. Relationship to World Models: the agent may construct an external task model composed of states, sub-goals, constraints, tools, and memory.

Predictive Artificial Intelligence Architectures — 12.3. Model-Based Reinforcement Learning

Model-based reinforcement learning learns or exploits a model of environmental dynamics. The agent can simulate the consequences of its actions before acting. See Sutton & Barto and Moerland et al..

Primary strength: efficient planning and anticipation. Main limitation: learning reliable models in complex environments remains difficult. Relationship to World Models: this is one of the most explicit forms of a World Model.

12.4. Model-Free Reinforcement Learning

Model-free reinforcement learning learns an action policy directly without constructing an explicit model of the environment. It has achieved remarkable success in games and several simulated environments. See Mnih et al. and Schulman et al..

Primary strength: strong performance in well-defined environments with clear reward functions. Main limitation: high training cost, limited data efficiency, and poor robustness outside the training distribution. Relationship to World Models: although it avoids an explicit World Model, it generally struggles with long-horizon planning and systematic generalization without predictive structure.

12.5. Imitation Learning and Learning from Demonstration

Imitation learning trains a system to reproduce observed behaviors. It plays a central role in robotics, autonomous driving, and software agents.

Primary strength: rapid learning from human demonstrations. Main limitation: behavior may be reproduced without deep understanding, leading to failures outside the training distribution. Relationship to World Models: demonstrations provide trajectories, but the agent often requires a predictive model to adapt to novel situations.

Predictive Artificial Intelligence Architectures — 12.6. Active Inference and the Free Energy Principle

Active Inference, associated with Friston, proposes that intelligent agents act to reduce uncertainty and minimize the discrepancy between predictions and observations. Policies are selected according to their expected ability to minimize free energy by jointly maximizing utility and information gain. See Friston et al. and de Vries.

Primary strength: a unified framework integrating perception, action, and uncertainty. Main limitation: considerable theoretical complexity and limited industrial adoption. Relationship to World Models: Active Inference relies on internal generative models and is therefore closely related to, rather than opposed to, World Models.

12.7. Causal Models and Probabilistic Reasoning

Causal Models seek to distinguish correlation from causation while enabling counterfactual reasoning: what would happen if a variable were changed? See Pearl and Schölkopf et al..

Primary strength: conceptual robustness and intervention capabilities. Main limitation: learning large-scale causal structures automatically remains extremely challenging. Relationship to World Models: a causal model is often an abstract World Model centered on causal mechanisms.

12.8. Neuromorphic and Brain-Inspired Architectures

Neuromorphic architectures investigate spiking neural networks, continuous plasticity, local memory, and highly energy-efficient computation.

Primary strength: biological inspiration and potentially high energy efficiency. Main limitation: lower technological maturity compared with mainstream deep learning architectures. Relationship to World Models: these architectures do not inherently provide a World Model, but they may constitute an effective substrate for continuous learning.

Predictive Artificial Intelligence Architectures — 12.9. Planning Through Search, MCTS, Programs, and Formal Verification

Planning may be performed through explicit search methods, including decision trees, Monte Carlo Tree Search, constraint solvers, theorem provers, and formal verification systems. See Kocsis & Szepesvári and Silver et al..

Primary strength: systematic exploration of alternative scenarios. Main limitation: combinatorial explosion and dependence on a formal representation of states. Relationship to World Models: every search tree assumes states and transitions and therefore relies on some form of World Model.

12.10. Evolutionary AI and Open-Ended Learning

Evolutionary AI seeks to generate increasingly complex behaviors through variation, selection, and open-ended environments. The objective is not merely to optimize a fixed task, but to encourage the emergence of novel capabilities.

Primary strength: open-ended exploration of behavioral diversity. Main limitation: computational cost, unpredictability, and limited controllability. Relationship to World Models: evolved agents may develop internal representations, although these are often difficult to interpret.

12.11. Metacognitive Architectures

Metacognitive architectures provide a system with self-assessment capabilities, enabling it to detect its own errors, estimate uncertainty, decide when to request assistance, verify hypotheses, or revise its strategy.

Primary strength: robustness, self-correction, and operational safety. Main limitation: objectively evaluating the quality of self-assessment remains difficult. Relationship to World Models: metacognition can supervise and regulate the use of a World Model, but it does not replace it.

13. Proposed Taxonomy of Predictive Artificial Intelligence Architectures

This taxonomy proposes seven dimensions for comparing candidate architectures capable of achieving robust artificial general intelligence.

1. Language: processing symbols, text, instructions, and dialogue.
2. Perception: learning from images, video, audio, sensors, or the surrounding environment.
3. Memory: storing, organizing, abstracting, and reusing experience.
4. Causality: distinguishing correlation, intervention, and consequence.
5. Action: operating within real, simulated, or software environments.
6. Prediction: anticipating future states and multiple possible scenarios.
7. Planning: selecting sequences of actions to achieve a goal.

This taxonomy deliberately avoids classifying approaches according to technological trends or implementation choices. Instead, it organizes them according to the cognitive functions they are required to perform.

The central question therefore becomes: Which architecture most effectively integrates these seven dimensions while ensuring robustness, safety, and verifiability?

Predictive Artificial Intelligence Architectures — 14. Comparative Matrix of Current Approaches

Qualitative scale: Low / Moderate / High / Very High.

This comparative matrix shows that World Models are not the only possible path toward advanced intelligence. However, nearly all advanced approaches ultimately face the same fundamental challenge: representing, predicting, remembering, acting, and planning.

15. Proposed Hybrid Architecture: LAMP-C

Epistemological status (Register A). Conceptual architecture · research framework · not experimentally validated at this stage.

To establish this dissertation as a foundation for future research, we propose a conceptual architecture named LAMP-C:

• L — Language: communication, instruction, and symbolic reasoning expressed through natural language.
• A — Abstraction: construction of hierarchical and compressed representations.
• M — Memory: storage, consolidation, forgetting, retrieval, and contradiction management through Agentic Memory and Experiential Memory.
• P — Prediction / Planning: simulation of possible futures and selection of appropriate actions.
• C — Causality / Control: intervention, counterfactual reasoning, verification, Runtime Integrity, and Cyber-Physical Trust.

Multimodal perception / data / language
                ↓
          Abstraction Encoder
                ↓
        Experiential Memory
                ↓
       Predictive World Model
                ↓
     Causal and Counterfactual Module
                ↓
 Planning Engine / Symbolic Engine / Tools
                ↓
Action: robot, API, software, or decision
                ↓
     Experience feedback and correction

This architecture is not intended as a finished technical product; it is a conceptual research framework. It provides a basis for comparing existing approaches while identifying the capabilities that remain absent from each of them.

LAMP-C is built upon a central hypothesis: advanced intelligence must be compositional. It does not emerge from a single monolithic model, but from the integration of language, perception, memory, prediction, causality, and control.

Predictive Artificial Intelligence Architectures — 16. Memory, Experience, and Cognitive Continuity

Without memory, an intelligent agent remains largely stateless. It may answer questions within a context window, yet it cannot develop continuity of experience.

Current AI systems are exploring several complementary forms of memory:

1. Contextual Memory: information available within the model’s context window.
2. Document Memory: retrieval of documents or passages through RAG.
3. Episodic Memory: records of interactions, actions, errors, and outcomes.
4. Semantic Memory: consolidated abstract knowledge.
5. Procedural Memory: strategies, methods, routines, and acquired skills.
6. Experiential Memory: action trajectories, feedback, failures, corrections, and accumulated learning.

Modern Tool-using Agents based on Large Language Models already investigate these memory mechanisms. See Du and Zhang et al..

Useful memory should do more than simply accumulate information. It must also filter, consolidate, forget, resolve contradictions, preserve confidentiality, and connect past experience to future decision-making.

A rigorous research program should therefore evaluate not only retrieval performance, but also whether memory genuinely improves decision quality.

17. Causality, Counterfactual Reasoning, and Robustness

Causality represents one of the major boundaries between statistical correlation and robust intelligence.

A statistical model may learn that two events are associated. A Causal Model seeks to understand what produces what. It enables questions such as:

• What would happen if I intervened on this variable?
• Does this action cause the observed effect, or merely reveal it?
• What would have happened if a different action had been taken?

Pearl formalized this distinction through causal and counterfactual reasoning. Schölkopf et al. further discuss the importance of causality for robust learning and out-of-distribution generalization.

A World Model without causality may capture only superficial regularities. Conversely, a Causal Model without perception may lack grounding in reality. A robust hybrid architecture should therefore integrate both.

Predictive Artificial Intelligence Architectures — 18. Scientific Evaluation of Candidate Architectures

To establish this dissertation as the foundation of a research program, its hypotheses must be scientifically falsifiable.

Predictive Artificial Intelligence Architectures — 18.1. Evaluation Framework

A candidate architecture should be evaluated according to ten dimensions:

1. Prediction: Can it accurately anticipate the evolution of an environment?
2. Counterfactual Reasoning: Can it simulate “What would happen if…” scenarios?
3. Planning: Can it select an effective sequence of actions?
4. Causality: Can it distinguish causal relationships from mere correlations?
5. Out-of-Distribution Robustness: Can it operate reliably in novel situations?
6. Long-Term Memory: Does it learn effectively from previous experiences?
7. Physical or Operational Grounding: Does it integrate language, perception, and action?
8. Explainability: Can its decisions be understood and analyzed?
9. Safety: Does it fail safely when necessary?
10. Trust Governance: Can its capabilities, access rights, and objectives be effectively controlled?

18.2. Falsifiable Hypotheses

Hypothesis H1. An architecture combining a Large Language Model, Experiential Memory, and a latent predictive model performs better on long-horizon planning tasks than a standalone LLM.

Hypothesis H2. Adding a Causal Model improves out-of-distribution robustness under changing environmental conditions.

Hypothesis H3. Consolidated Experiential Memory reduces the recurrence of errors across multi-session tasks.

Hypothesis H4. A Neuro-symbolic AI architecture reduces hallucinations in tasks involving formal constraints.

Hypothesis H5. Latent World Models predict the consequences of physical actions more accurately than purely text-based models.

18.3. Candidate Experimental Protocols

• Simulated robotic or physics-based environments.
• Multi-step planning tasks involving hidden constraints.
• Multi-session memory benchmarks.
• Causal and counterfactual reasoning benchmarks.
• Out-of-distribution evaluation scenarios.
• Formal verification of generated plans.
• Comparative evaluation of standalone LLMs, Tool-using Agents, LLMs with memory, LLMs with World Models, and the proposed LAMP-C architecture.

19. Mapping the Scientific Debates

A reference document should present scientific disagreements as well as defend its own thesis.

Predictive Artificial Intelligence Architectures — 19.1. Is Text Alone Sufficient?

Some researchers argue that scale, data, and external tools will ultimately enable Large Language Models to build sufficiently rich internal representations. Others contend that text alone cannot provide the grounding required for causal and physically situated intelligence.

19.2. Do Large Language Models Truly Reason?

LLMs sometimes produce reasoning that is useful and convincing. However, distinguishing robust reasoning from the imitation of common reasoning patterns or implicit search within textual representations remains an open scientific question.

19.3. Can Causality Emerge from Scale?

Causal relationships may be learned partially from data, but intervention and counterfactual reasoning often require additional representational structures beyond statistical scaling alone.

Predictive Artificial Intelligence Architectures — 19.4. Is Physical Embodiment Necessary?

Artificial intelligence can clearly be useful without a physical body. However, intelligence approaching that of humans or animals may require some form of embodied experience, whether real or simulated.

19.5. Are Video Models Sufficient?

Video models learn visual dynamics effectively, yet they may still lack causal understanding, intentionality, hidden physical constraints, and validation through interaction with the real world.

19.6. Is Neuro-symbolic AI a Transitional Stage or a Long-Term Direction?

Neuro-symbolic AI may serve either as a reasoning and control layer or as a central component of future hybrid architectures.

Predictive Artificial Intelligence Architectures — 19.7. Are Tool-using Agents a Sustainable Architecture?

They are already highly valuable in industrial applications, but their long-term robustness depends heavily on memory, external tools, verification mechanisms, and effective control.

Predictive Artificial Intelligence Architectures — 20. Proposed Research Program

20.1. Overall Objective

Design and evaluate a hybrid architecture capable of integrating language, perception, memory, prediction, causality, and planning.

20.2. Year 1: Taxonomy and Experimental Foundation

• Finalize the proposed taxonomy.
• Develop the comparative evaluation matrix.
• Select appropriate benchmark suites.
• Develop an initial prototype combining an LLM, memory, and external tools.
• Evaluate the limitations of standalone LLMs on planning tasks.

Predictive Artificial Intelligence Architectures — 20.3. Year 2: Memory, Causality, and Latent World Models

• Integrate Experiential Memory.
• Add a causal or counterfactual reasoning module.
• Evaluate a latent predictive model within a simulated environment.
• Compare model-free RL, model-based RL, and Tool-using Agents.

20.4. Year 3: LAMP-C Architecture and Validation

• Integrate language, abstraction, memory, prediction, and causality.
• Evaluate out-of-distribution robustness.
• Measure reductions in repeated errors.
• Assess safety and explainability.
• Publish the framework, experimental results, and identified limitations.

20.5. Scientific Deliverables

• Position paper.
• Comparative survey in French and English.
• LAMP-C taxonomy.
• Internal benchmark for planning and memory.
• Experimental prototype.
• Evaluation report.
• Continuously maintained annotated bibliography.

21. Risks, Trust Governance, and Safety

Advanced AI architectures introduce specific risks.

A World Model improves planning capabilities, yet more effective planning may also increase a system’s ability to pursue unintended objectives. Persistent memory strengthens continuity but raises important issues regarding confidentiality, the right to be forgotten, and the persistence of erroneous knowledge. External tools improve operational effectiveness but also introduce risks associated with uncontrolled execution.

Trust Governance should therefore be integrated into the architecture from the outset:

• capability control;
• logging and traceability;
• plan verification;
• operational action limits;
• clear separation between prediction, decision, and execution;
• memory management;
• explainability;
• auditability;
• safe failure (fail-safe);
• goal alignment.

Consequently, any research program on Predictive Artificial Intelligence Architectures must also be conceived as a research program in AI safety and Cyber-Physical Trust.

Predictive Artificial Intelligence Architectures — 22. A Defensible Scientific Position

This dissertation does not claim to demonstrate that World Models constitute the only path toward artificial general intelligence. It defends a broader and more robust position: any architecture aiming to achieve reliable intelligence, planning capabilities, and generalization will need to include, explicitly or implicitly, predictive, memory-based, causal, and actionable capabilities.

This position avoids two extremes. The first would be to reduce Large Language Models to simple systems with no internal representations at all: work such as Gurnee & Tegmark 2023 suggests that they can encode certain spatial and temporal reference structures. The second would be to conclude that text alone is sufficient to produce robust embodied intelligence: limitations such as the Reversal Curse, the absence of direct sensorimotor grounding, and weaknesses in planning indicate that such a conclusion remains fragile.

The defensible thesis is therefore the following:

This formulation makes the dissertation compatible with competing and complementary approaches, including Neuro-symbolic AI, Tool-using Agents, RAG, Active Inference, causality, embodied robotics, reinforcement learning, and hybrid architectures. It also supports the argument that World Models are less a doctrine than a remarkable instance of a broader cognitive function: anticipating what may happen given the current state and the possible actions. See Craik 1943, Johnson-Laird 1983, Sutton & Barto 2018, Ha & Schmidhuber 2018, and LeCun 2022.

23. State of the Art at the Time of Writing: Research, Industrialization, and Observed Results

State of the art documented up to 2026-07-07; this field is evolving rapidly. This section distinguishes three levels:

1. scientific research: articles, surveys, benchmarks, and experimental architectures;
2. industrialized implementation: products, platforms, standards, regulations, or already deployed uses;
3. observed results: measured benefits, real limitations, disappointing outcomes, or persistent risks.

The objective is not to provide an exhaustive list of AI products, but to position Predictive AI Architectures within their operational reality: what already works, what is progressing, what remains fragile, and what still needs to be demonstrated.

Predictive Artificial Intelligence Architectures — 23.1. Short Synthesis

At the time of writing, the state of the art shows a clear convergence: the most effective systems do not rely on a single component. They generally combine a language model, memory or external retrieval, tools, guardrails, access policies, evaluations, and sometimes specialized modules for vision, planning, cybersecurity, or robotics.

Industrialized LLMs are already effective for writing assistance, code generation, user support, document analysis, augmented search, and support for security teams. However, their limitations remain well documented: hallucinations, context dependence, fragile long-horizon planning, agent security, variable quality of generated code, data leakage risks, and the continuing need for supervision.

World Models and predictive video models are progressing rapidly in research, particularly with V-JEPA 2 and the 2025–2026 surveys on robotics and embodied AI. However, their full industrialization remains limited: results are promising for video understanding, prediction, zero-shot planning, and controlled robotics, but they are not yet equivalent to open-world autonomous general intelligence.

Cybersecurity and Digital Identity approaches are the most industrialized from a normative standpoint: NIST SP 800-63-4, OWASP LLM Top 10, NIST AI RMF, NIST CSF 2.0, ETSI EN 303 645, the Cyber Resilience Act, and the EU AI Act already form a reference foundation. WebAuthn/FIDO and Passkeys may also be cited as external points of comparison for passwordless authentication, without constituting the Freemindtronic trust foundation. The real outcome is clear: digital trust is evolving toward strong identity, security by design, AI risk governance, and phishing resistance. However, the integration of AI, identity, connected objects, and cyber-physical safety remains an emerging field of applied research.

23.2. LLMs and Tool-using Agents: Strong Industrialization, Still Incomplete Robustness

Large Language Models are the most industrialized building blocks of contemporary AI. They are now integrated into office environments, search engines, development platforms, support tools, business assistants, augmented SOCs, and document workflows.

Examples of Already Industrialized Implementations

Predictive Artificial Intelligence Architectures — 24. Benchmarks and Evaluation Protocols

A reference dissertation should propose not only concepts but also evaluation criteria. A candidate Predictive Artificial Intelligence Architecture must be assessed through protocols that measure its ability to predict, plan, remember, act, explain its decisions, and fail safely.

24.1. Evaluating Prediction

Key questions:

• Can the system accurately predict the evolution of an environment?
• Can it represent multiple possible futures?
• Does it distinguish epistemic uncertainty from aleatoric uncertainty?
• Does it predict in pixel space, token space, or an abstract latent representation?

Relevant references: Ha & Schmidhuber 2018, Moerland et al. 2023, Bardes et al. 2024, Assran et al. 2025.

Predictive Artificial Intelligence Architectures — 24.2. Evaluating Planning

Key questions:

• Can the system decompose a complex task into manageable steps?
• Can it compare multiple alternative plans?
• Can it revise a plan after failure?
• Can it plan under temporal, energy, or regulatory constraints?

Relevant references: Kocsis & Szepesvári 2006, Silver et al. 2018, Huang et al. 2024, ReAct.

24.3. Evaluating Memory

Key questions:

• Does the system retain relevant episodes?
• Can it consolidate experience into abstract knowledge?
• Can it forget information that is unnecessary or potentially harmful?
• Can it manage contradictions, corrections, and the right to be forgotten?

Relevant references: Zhang et al. 2024, Du 2026, Lewis et al. 2020.

24.4. Evaluating Causality and Counterfactual Reasoning

Key questions:

• Can the system distinguish correlation from causation?
• Can it answer “What would happen if…?” questions?
• Can it identify the relevant intervention variables?
• Does it remain robust under distribution shifts?

Relevant references: Pearl 2009, Schölkopf et al. 2021, Lake et al. 2017.

Predictive Artificial Intelligence Architectures — 24.5. Evaluating Out-of-Distribution Robustness

Key questions:

• Can the system generalize to previously unseen scenes, objects, or rules?
• Can it recognize its own limitations?
• Can it suspend an action instead of producing a plausible but incorrect response?

Relevant references: Berglund et al. 2023, Bender et al. 2021, World Model for Robot Learning 2026.

24.6. Evaluating Trust Governance

Key questions:

• Are generated plans auditable?
• Is memory fully traceable?
• Are actions clearly separated from decisions?
• Are guardrails, uncertainty thresholds, and fail-safe mechanisms implemented?

A comprehensive benchmark should therefore combine prediction tasks, planning tasks, long-term memory tasks, causal reasoning tasks, out-of-distribution evaluations, decision auditing, and safety testing.

25. Agentic Memory: The Missing Link

Memory is often treated as a secondary module. This is a fundamental mistake. Without memory, an intelligent agent possesses no continuity of experience. Without continuity, it cannot learn sustainably from its actions, correct recurring mistakes, manage contradictions, or build a stable functional identity.

A World Model without Experiential Memory risks remaining only a local prediction mechanism. To become a cumulative intelligence, it must be coupled with a memory system capable of preserving experience, extracting abstractions, forgetting irrelevant details, managing contradictions, and reusing acquired knowledge in new situations.

25.1. Three Levels of Memory

1. Contextual Memory: the information contained within the model’s current context window.
2. External Memory: documents, vector databases, RAG systems, logs, and knowledge graphs.
3. Experiential Memory: episodes, errors, decisions, consequences, abstraction, consolidation, and forgetting.

Predictive Artificial Intelligence Architectures — 25.2. The Write–Manage–Read Loop

Recent research formalizes agent memory as a continuous cycle:

Observation / action
        ↓
Memory writing
        ↓
Memory management:
compression, hierarchy, contradiction handling, forgetting
        ↓
Selective retrieval
        ↓
Decision / planning
        ↓
New action

This loop must be integrated with perception, action, access control, and data Trust Governance. See Du 2026 and Zhang et al. 2024.

25.3. Memory and Operational Sovereignty

Agentic Memory also introduces sovereignty requirements, including data localization, encryption, traceability, the right to be forgotten, human oversight, separation between personal and professional memories, and protection against memory poisoning.

Memory is therefore not merely a technical challenge; it is fundamentally a matter of Trust Governance.

Predictive Artificial Intelligence Architectures — 26. AI-TRL Maturity Framework

To transform this dissertation into the foundation of a research program, the maturity of candidate architectures must be assessed. The following framework adapts the spirit of Technology Readiness Levels (TRLs) to Predictive Artificial Intelligence Architectures.

This framework enables meaningful comparisons between approaches without conflating them. A Large Language Model may score very highly on language while remaining limited in embodied interaction. A World Model may excel at prediction while remaining weak in Trust Governance. A hybrid architecture should therefore strive for balanced progress across all dimensions.

27. Manifesto for Predictive, Memory-Driven, and Governable AI

1. Language is not the world. Text describes reality, but it cannot replace sensory experience, action, or causality.
2. Predicting tokens is not the same as predicting consequences. An intelligent system that acts must anticipate the effects of its actions.
3. Memory is not merely a document repository. It should become a continuity of experience through consolidation, forgetting, and controlled contradiction management.
4. Causality cannot be reduced to correlation. A robust AI system must reason about interventions and counterfactuals.
5. Planning requires simulatable futures. Choosing an action presupposes comparing multiple possible trajectories.
6. Action requires safety control. The greater a system’s capacity to act, the greater the need for Trust Governance, auditability, and operational constraints.
7. Abstraction is prediction-oriented compression. Irrelevant details must be discarded while preserving the variables that matter for prediction.
8. Artificial general intelligence will most likely be hybrid. Language, perception, Agentic Memory, Causal Models, tools, and latent World Models will need to cooperate.
9. Evaluation must consider long-term behavior and out-of-distribution performance. Short benchmarks alone cannot adequately measure robustness.
10. A powerful AI system must know how to fail safely. Refusing, suspending action, requesting verification, or limiting execution may be more intelligent than producing a plausible but incorrect response.

This manifesto summarizes the central ambition of this dissertation: to move beyond generative AI centered on text production toward Predictive Artificial Intelligence Architectures that integrate prediction, Agentic Memory, causality, action, and Trust Governance.

Predictive Artificial Intelligence Architectures — 28. Appendix: Doctoral Research Project / Consortium

28.1. Possible Title

Toward a Hybrid Architecture for Predictive Intelligence: Memory, Causality, World Models, and Tool-using Agents.

Predictive Artificial Intelligence Architectures — 28.2. Research Problem

Current AI architectures excel at language generation. However, they remain fragile when they need to act over time, retain experience, generalize beyond the training distribution, reason causally, and plan within open environments.

This research project aims to study whether a hybrid architecture combining a Large Language Model, a World Model, Agentic Memory, Causal Models, and symbolic control can improve the robustness and governability of autonomous agents.

28.3. Research Hypotheses

• H1: structured Experiential Memory reduces repeated errors in LLM-based agents.
• H2: a latent predictive model improves planning compared with purely text-based planning.
• H3: adding a Causal Model improves out-of-distribution robustness.
• H4: Neuro-symbolic AI control reduces incoherent or forbidden actions.
• H5: a hybrid LAMP-C architecture achieves greater governability than a Tool-using LLM Agent alone.

28.4. Scientific Bottlenecks

• Learning the right abstractions without reconstructing every detail.
• Combining long-term memory with confidentiality requirements.
• Evaluating causality and counterfactual reasoning.
• Controlling action in open environments.
• Preventing memory poisoning.
• Maintaining auditability despite opaque neural modules.

Predictive Artificial Intelligence Architectures — 28.5. Methodology

1. Conduct a structured literature review.
2. Define internal benchmarks for memory, planning, causality, and safety.
3. Develop an agentic prototype combining an LLM, RAG, memory, a simulator, and a symbolic verifier.
4. Progressively integrate a latent predictive model.
5. Evaluate the system against a standalone LLM, a RAG agent, a Tool-using Agent, an agent with memory, and a hybrid agent.
6. Analyze failures, including hallucination, causal errors, impossible plans, and contradictory memory.
7. Publish the results, limitations, and evaluation protocols.

28.6. Deliverables Over 36 Months

28.7. Potential Applications

• Robotics and embodied agents.
• Long-term professional assistants.
• AI Cybersecurity and incident analysis.
• Governed critical systems.
• Sovereign off-cloud agents.
• Decision support under regulatory constraints.

Predictive Artificial Intelligence Architectures — 28.8. Success Criteria

• Measurable reduction in repeated errors.
• Improved planning under constraints.
• Greater out-of-distribution robustness.
• Complete logging of decisions and actions.
• Explicit control over action capabilities.
• Reproducibility of evaluation protocols.

Predictive Artificial Intelligence Architectures — Limitations, Falsifiability, and Scope of Validity

This section consolidates, for Freemindtronic’s public reference publication, material that is distributed elsewhere in the dissertation (§11.5, §18, §19, Appendix A.6). Its purpose is to make the document defensible for a skeptical reader, whether a researcher, auditor, journalist, or industrial partner.

Predictive Artificial Intelligence Architectures — What This Dissertation Does Not Claim to Prove

This document is not:

• an exhaustive PRISMA-style systematic review;
• an independent security audit or a compliance attestation (eIDAS, Common Criteria, FIPS, etc.);
• a published quantitative benchmark comparing EviSKMS with FIDO, PKI, or competing solutions;
• an enabling technical disclosure allowing reproduction of Gen2 mechanisms or post-patent extensions;
• peer validation in the strict sense of publication in a peer-reviewed journal.

It is: an interdisciplinary framework for Predictive Artificial Intelligence Architectures; an applied positioning in cybersecurity and Cyber-Physical Trust (§29.1–§29.13); and a bridge to the companion DNA/EviDNA dissertation for cryptographic details and state-of-the-art comparisons.

Scope of Validity by Register

Falsifiable Hypotheses — Predictive Artificial Intelligence Dimension

Hypotheses H1 to H5 in §18.2 concern LAMP-C and hybrid architectures integrating memory, causality, World Models, and Neuro-symbolic AI. They remain valid for the AI research dimension of this dissertation.

However, their confirmation or refutation requires the experimental protocols described in §18.3 and §24.

Predictive Artificial Intelligence Architectures — Falsifiable Hypotheses — Digital Trust Dimension (EviSKMS Gen1)

Hypotheses H-C1 to H-C5 concerning continuity, fail-closed behavior, DDNA, anti-replay mechanisms, and differentiation from existing standards are formulated and detailed in the companion DNA/EviDNA dissertation.

Global Refutation Conditions for the Freemindtronic Positioning

The framework defended in this dissertation would be significantly weakened if one of the following conditions were established publicly and reproducibly:

1. Gen2 presented without register qualification, despite its detailed mechanisms belonging to Register C.
2. Systemic bypass of fail-closed, RI, or DRT continuity controls within the qualified sovereign-local scope, without a documented corrective measure.
3. Absence of correlation between the patented segmentation mechanism and the industrialized Gen1 mechanisms, indicating a technical or documentary discontinuity.
4. Independent benchmark evidence showing that properly deployed MFA/WebAuthn achieves the same temporal continuity and runtime governance properties without an additional layer, across the same adversarial scenarios.
5. Unintentional enabling disclosure in public communications, including the dissertation, videos, or press releases, allowing a third party to reproduce Gen2 or post-patent extensions.

Methodological Constraint Related to Intellectual Property

The controlled publication strategy based on Registers A / B / C strengthens IP protection. However, it also reduces immediate external falsifiability: a third party cannot reproduce or deeply audit mechanisms classified as Register C without an agreement.

This constraint is intentional. It requires a clear distinction between:

• what is publicly verifiable, including product existence, declared automated tests, granted patents, and timestamped disclosures;
• what is verifiable under NDA (Register B);
• what is deliberately not published (Register C).

Full scientific recognition will require third-party evaluations on authorized scopes, after IP protection has been secured, in accordance with §1.2 of the companion dissertation.

Predictive Artificial Intelligence Architectures — Epistemological Modesty

This dissertation adopts the position of an inventor-researcher: field observation and industrialization provide strong signals, but they do not replace independent validation.

General Conclusion

Large Language Models have demonstrated the power of large-scale statistical learning. They will remain an essential component of modern artificial intelligence, because language is the primary medium of explicit human knowledge.

However, language alone is likely insufficient to produce robust artificial general intelligence. An intelligence capable of action must retain experience, represent context, anticipate the consequences of its actions, reason causally, plan, and control its own limits.

World Models represent a major path toward this capability, but they are not the only one. Neuro-symbolic AI, Tool-using Agents, RAG, persistent memory, reinforcement learning, Active Inference, Causal Models, search-based planning, and embodied architectures each contribute part of the solution.

The central contribution of this dissertation is to shift the analytical focus toward a broader framework: Predictive Artificial Intelligence Architectures. Within this framework, World Models are no longer the school of thought to defend; rather, they become one pillar of a larger architecture grounded in memory, abstraction, causality, action, and Trust Governance.

The applied cybersecurity dimension shows why this approach is becoming critical. Identity, context, memory, behavior, proof, action, and consequence must be connected in order to maintain Trust Continuity between humans, AI agents, machines, and connected devices.

The Cryptographic Genome / EviDNA trajectory, including CryptPeer/EviSKMS industrialization, illustrates this evolution on the sovereign trust side. It is developed in the companion DNA/EviDNA dissertation. The section on limitations and falsifiability specifies the scope of validity of the present document.

A major architectural evolution in AI will likely not consist merely of a larger model. Instead, it will depend on a better-structured architecture: language, abstraction, memory, prediction, causality, action, and control, subject to the methodological limits explicitly stated in this dissertation.

Predictive Artificial Intelligence Architectures — Annotated Scientific Bibliography

This bibliography is designed as an interactive section. Each entry includes a stable internal link, one or more official or primary links, and an indication of how it is used in the dissertation.

Quick Bibliography Index

• Cognitive origins and symbolic grounding
• Limitations and capabilities of LLMs
• Cognitive science and human learning
• Human vision and sensory flow
• Reinforcement learning and World Models
• JEPA, video, and embodied robotics
• RAG, tools, agents, and memory
• Industrialized implementations and observed real-world outcomes
• Neuro-symbolic AI
• Active Inference
• Causality
• Model-free RL, MCTS, and AlphaZero
• Cybersecurity, Digital Identity, IoT, and safety

Cognitive Origins and Symbolic Grounding

Craik, K. J. W. (1943). The Nature of Explanation.

Official / primary links: PhilPapers · Google Books / CUP Archive · Internet Archive. A foundational reference introducing the concept of an internal small-scale model of reality. It is useful for showing that the idea of a World Model predates modern artificial intelligence. Use in this dissertation: historical origins of internal models, mental simulation, and prediction before action. ↩ Back to the bibliography index

Johnson-Laird, P. N. (1983). Mental Models.

Official / primary links: Google Books / Harvard University Press · ACM Guide. A landmark work in cognitive psychology introducing the theory of mental models. It establishes a conceptual bridge between human reasoning and the internal simulation of possible situations. Use in this dissertation: cognition, internal simulation, and reasoning about possible situations. ↩ Back to the bibliography index

Harnad, S. (1990). The Symbol Grounding Problem.

Official / primary links: Oxford Computer Science PDF. A classic paper addressing the challenge of assigning meaning to symbols that are connected only to other symbols. Use in this dissertation: symbolic grounding and the limitations of language in the absence of perception and action. ↩ Back to the bibliography index

Large Language Models (LLMs): Capabilities and Limitations

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?

Official / primary links: ACM Digital Library · ACM FAccT 2021. A landmark paper arguing that scaling language models does not, by itself, produce grounded understanding. The authors examine issues related to data quality, bias, environmental cost, and the limitations of purely statistical language modeling. Use in this dissertation: critical analysis of LLM limitations, symbolic grounding, and the distinction between language generation and genuine understanding. ↩ Back to the bibliography index

Gurnee, W., & Tegmark, M. (2023). Language Models Represent Space and Time.

Official / primary links: arXiv. This study provides evidence that Large Language Models develop internal representations encoding spatial and temporal relationships, suggesting that statistical learning can produce latent predictive representations extending beyond surface-level language patterns. Use in this dissertation: internal representations in LLMs, latent world representations, and the scientific debate on emergent cognitive capabilities. ↩ Back to the bibliography index

Berglund, L., et al. (2023). The Reversal Curse: LLMs Trained on “A is B” Fail to Learn “B is A”.

Official / primary links: arXiv. This paper introduces the Reversal Curse, demonstrating that language models may learn directional relationships without reliably inferring their inverse. The findings highlight limitations in generalization and relational reasoning. Use in this dissertation: limits of generalization, causal reasoning, and the robustness of internal representations. ↩ Back to the bibliography index

Brown, T. B., et al. (2020). Language Models are Few-Shot Learners.

Official / primary links: arXiv. The foundational GPT-3 paper demonstrating that scaling transformer-based language models enables strong few-shot, one-shot, and zero-shot performance across a wide range of natural language tasks. Use in this dissertation: emergence of in-context learning, scaling laws, and the practical capabilities of Large Language Models. ↩ Back to the bibliography index

OpenAI (2023). GPT-4 Technical Report.

Official / primary links: arXiv. Describes the architecture, evaluation methodology, capabilities, and limitations of GPT-4, including benchmark performance and safety considerations. Use in this dissertation: state of the art in industrial LLMs, evaluation methodology, and current operational capabilities. ↩ Back to the bibliography index

Cognitive Science and Human Learning

Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building Machines That Learn and Think Like People.

Official / primary links: Science. A seminal article arguing that human-level intelligence requires more than statistical pattern matching. The authors advocate integrating compositionality, causality, intuitive physics, intuitive psychology, and efficient learning inspired by cognitive science. Use in this dissertation: cognitive foundations of Predictive Artificial Intelligence Architectures, compositional representations, and causal reasoning. ↩ Back to the bibliography index

Dehaene, S. (2020). How We Learn: Why Brains Learn Better Than Any Machine… for Now.

Official / primary links: Penguin Random House · Collège de France. Dehaene explores the principles underlying human learning, emphasizing attention, active engagement, error correction, abstraction, and consolidation. Use in this dissertation: human learning mechanisms, abstraction, Agentic Memory, and the relationship between learning and prediction. ↩ Back to the bibliography index

Friston, K. (2010). The Free-Energy Principle: A Unified Brain Theory?

Official / primary links: Nature Reviews Neuroscience. This influential paper proposes that biological systems minimize variational free energy through prediction and continual interaction with their environment, providing a theoretical foundation for perception, action, and learning. Use in this dissertation: predictive cognition, Active Inference, and theoretical foundations of predictive intelligence. ↩ Back to the bibliography index

Clark, A. (2013). Whatever Next? Predictive Brains, Situated Agents, and the Future of Cognitive Science.

Official / primary links: Cambridge University Press. Clark argues that cognition is fundamentally predictive, with the brain continuously generating and updating models of the world through perception and action. Use in this dissertation: predictive cognition, World Models, Active Inference, and predictive representations. ↩ Back to the bibliography index

Human Vision and Sensory Processing

Gibson, J. J. (1979). The Ecological Approach to Visual Perception.

Official / primary links: Google Books. Gibson introduced the ecological theory of perception, arguing that perception is fundamentally rooted in direct interaction with the environment rather than in passive image processing. Use in this dissertation: perceptual grounding, embodied intelligence, and the relationship between perception and action. ↩ Back to the bibliography index

Marr, D. (1982). Vision: A Computational Investigation into the Human Representation and Processing of Visual Information.

Official / primary links: MIT Press. A foundational work in computational vision proposing a hierarchical theory of visual processing, from low-level sensory signals to abstract representations. Use in this dissertation: hierarchical representations, abstraction, predictive perception, and World Models. ↩ Back to the bibliography index

DiCarlo, J. J., Zoccolan, D., & Rust, N. C. (2012). How Does the Brain Solve Visual Object Recognition?

Official / primary links: Neuron. This review explains how the primate visual system develops invariant object representations while preserving behaviorally relevant information. Use in this dissertation: perceptual representations, abstraction, and biologically inspired approaches to Predictive Artificial Intelligence Architectures. ↩ Back to the bibliography index

Reinforcement Learning and World Models

Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2nd ed.).

Official / primary links: Official online edition. The reference textbook on reinforcement learning, covering value functions, policy optimization, temporal-difference learning, planning, model-based reinforcement learning, and exploration. Use in this dissertation: reinforcement learning, planning, predictive decision-making, and the foundations of World Models. ↩ Back to the bibliography index

Ha, D., & Schmidhuber, J. (2018). World Models.

Official / primary links: arXiv · Official project website. This pioneering work demonstrates how an agent can learn a compact latent model of its environment and use it for simulation, prediction, and planning before acting. Use in this dissertation: foundational reference for modern World Models, latent predictive representations, and predictive planning. ↩ Back to the bibliography index

Moerland, T. M., Broekens, J., & Jonker, C. M. (2023). Model-Based Reinforcement Learning: A Survey.

Official / primary links: arXiv. A comprehensive survey of model-based reinforcement learning, covering predictive environment models, planning algorithms, uncertainty estimation, and sample-efficient learning. Use in this dissertation: state of the art in model-based reinforcement learning, predictive planning, and World Models. ↩ Back to the bibliography index

Hafner, D., Pasukonis, J., Ba, J., & Lillicrap, T. (2023). Mastering Diverse Domains through World Models.

Official / primary links: arXiv. This paper presents DreamerV3, showing that a single World Model architecture can learn efficiently across a wide range of environments using latent imagination and predictive planning. Use in this dissertation: scalable World Models, latent imagination, generalization, and planning across heterogeneous environments. ↩ Back to the bibliography index

Silver, D., Hubert, T., Schrittwieser, J., et al. (2018). A General Reinforcement Learning Algorithm That Masters Chess, Shogi, and Go through Self-Play.

Official / primary links: Science. The AlphaZero paper demonstrates how planning, self-play, and predictive search can achieve superhuman performance without handcrafted domain knowledge. Use in this dissertation: planning, predictive search, reinforcement learning, and model-guided decision-making. ↩ Back to the bibliography index

Kocsis, L., & Szepesvári, C. (2006). Bandit Based Monte-Carlo Planning.

Official / primary links: Springer. This paper introduces Monte Carlo Tree Search (MCTS) with UCT, a breakthrough planning algorithm that balances exploration and exploitation through predictive simulation. Use in this dissertation: planning, predictive search, decision-making, and simulation-based reasoning. ↩ Back to the bibliography index

JEPA, Video, and Embodied Robotics

Bardes, A. et al. (2024). JEPA / V-JEPA Works.

Official / primary links: arXiv — Revisiting Feature Prediction for Learning Visual Representations from Video. A reference on learning predictive representations in latent space. Use in this dissertation: explaining why predicting abstract representations may be preferable to reconstructing every pixel. ↩ Back to the bibliography index

Assran, M. et al. (2025). V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning.

Official / primary links: arXiv · Meta AI — V-JEPA 2. Useful for discussing video prediction, abstract representations, and physical planning. Use in this dissertation: linking video, physical understanding, prediction, and planning. ↩ Back to the bibliography index

World Model for Robot Learning: A Comprehensive Survey (2026).

Official / primary links: arXiv · arXiv HTML. A recent survey on World Models in robotics, covering their paradigms, applications, limitations, and relationship to planning. Use in this dissertation: 2025–2026 state of the art, embodied robotics, benchmarks, and future research directions. ↩ Back to the bibliography index

A Comprehensive Survey on World Models for Embodied AI (2025).

Official / primary links: arXiv. A survey on World Models for embodied AI. Use in this dissertation: Appendix A.3, robotics, simulation, and embodied AI. ↩ Back to the bibliography index

RAG, Tool-Using Agents, and Agentic Memory

Lewis, P., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.

Official / primary links: arXiv. This foundational paper introduced Retrieval-Augmented Generation (RAG), combining Large Language Models with external knowledge retrieval to improve factual accuracy and reduce hallucinations. Use in this dissertation: retrieval-augmented reasoning, external knowledge integration, and the limits of document retrieval as a substitute for understanding. ↩ Back to the bibliography index

Yao, S., et al. (2023). ReAct: Synergizing Reasoning and Acting in Language Models.

Official / primary links: arXiv. ReAct demonstrates how interleaving reasoning traces with actions enables Large Language Models to interact more effectively with external tools and environments. Use in this dissertation: Tool-using Agents, reasoning-action loops, planning, and agent orchestration. ↩ Back to the bibliography index

Schick, T., et al. (2023). Toolformer: Language Models Can Teach Themselves to Use Tools.

Official / primary links: arXiv. Toolformer shows that Large Language Models can learn autonomously when and how to invoke external tools during inference, improving task performance without explicit supervision. Use in this dissertation: Tool-using Agents, autonomous tool selection, and hybrid Predictive Artificial Intelligence Architectures. ↩ Back to the bibliography index

Zhang, Y., et al. (2024). A Survey of Memory Mechanisms for Large Language Model Agents.

Official / primary links: arXiv. This survey reviews memory architectures for LLM-based agents, including memory writing, retrieval, consolidation, forgetting, contradiction handling, and long-term knowledge management. Use in this dissertation: Agentic Memory, Experiential Memory, long-term memory architectures, and trust-aware memory management. ↩ Back to the bibliography index

Du, X. (2026). Large Language Model Agent Memory: A Survey.

Official / primary links: arXiv. A comprehensive survey examining memory models for autonomous AI agents, including memory organization, lifecycle management, retrieval strategies, governance, evaluation, and future research directions. Use in this dissertation: Agentic Memory, Experiential Memory, memory governance, and persistent AI agents. ↩ Back to the bibliography index

Predictive Artificial Intelligence Architectures — Limitations and Capabilities of LLMs

Bender, E. M., Gebru, T., McMillan-Major, A., & Mitchell, M. (2021). On the Dangers of Stochastic Parrots.

Official / primary links: ACM DOI · Author PDF. An influential critique of Large Language Models, useful for addressing risks, grounding, bias, and the limits of text-only learning. Use in this dissertation: scientific caution regarding LLMs, scaling risks, and limits of understanding. ↩ Back to the bibliography index

Gurnee, W., & Tegmark, M. (2023). Language Models Represent Space and Time.

Official / primary links: arXiv · Official code. An important reference for qualifying critiques of LLMs: some models appear to encode spatial and temporal representations. Use in this dissertation: acknowledging that LLMs may contain fragments of World Models. ↩ Back to the bibliography index

Berglund, L. et al. (2023). The Reversal Curse.

Official / primary links: arXiv · OpenReview PDF. This work demonstrates a weakness in the relational generalization of autoregressive LLMs. Use in this dissertation: limits of relational reasoning and inverse generalization. ↩ Back to the bibliography index

Cognitive Science and Human Learning

Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building Machines That Learn and Think Like People.

Official / primary links: arXiv · PubMed · Stanford PDF. A major reference in cognitive science for Causal Models, intuitive physics, intuitive psychology, and rapid learning. Use in this dissertation: central argument for moving beyond purely text-based learning. ↩ Back to the bibliography index

Human Vision and Sensory Flow

Koch, K. et al. (2006). How Much the Eye Tells the Brain.

Official / primary links: PMC / NIH · EurekAlert / Penn. This work is useful for cautiously framing comparisons between human visual flow and the textual data processed by LLMs. The order of magnitude of retinal information transmission should be treated carefully; these estimates must not be presented as a strict equivalence between human vision and textual tokens. Use in this dissertation: cautious formulation of the passage on the four-year-old child. ↩ Back to the bibliography index

Predictive Artificial Intelligence Architectures — Reinforcement Learning and World Models

Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction.

Official / primary links: Official book website · Stanford PDF. A central reference on reinforcement learning, especially the distinction between model-based and model-free methods. Use in this dissertation: foundation for the distinction between action, reward, environment models, and planning. ↩ Back to the bibliography index

Moerland, T. M., Broekens, J., Plaat, A., & Jonker, C. M. (2023). Model-Based Reinforcement Learning: A Survey.

Official / primary links: ACM / Foundations and Trends · arXiv. This survey is useful for positioning model-based reinforcement learning as an approach to planning and anticipation. Use in this dissertation: integration of learning, environmental dynamics, and planning. ↩ Back to the bibliography index

Ha, D., & Schmidhuber, J. (2018). World Models.

Official / primary links: arXiv · Official interactive website. A modern explicit reference on World Models in AI: compressed representation, latent dynamics, and an agent trained within an internal model. Use in this dissertation: modern definition of World Models. ↩ Back to the bibliography index

LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence.

Official / primary links: OpenReview PDF. A structuring position on the limits of LLMs alone and the need for World Models, memory, perception, and planning. Use in this dissertation: autonomous architecture, latent-space prediction, and the role of memory and action. ↩ Back to the bibliography index

JEPA, Video, and Embodied Robotics

Bardes, A. et al. (2024). JEPA / V-JEPA Works.

Official / primary links: arXiv — Revisiting Feature Prediction for Learning Visual Representations from Video. A reference on learning predictive representations in latent space. Use in this dissertation: explaining why predicting abstract representations may be preferable to reconstructing every pixel. ↩ Back to the bibliography index

Assran, M. et al. (2025). V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning.

Official / primary links: arXiv · Meta AI — V-JEPA 2. Useful for discussing video prediction, abstract representations, and physical planning. Use in this dissertation: connecting video, physical understanding, prediction, and planning. ↩ Back to the bibliography index

World Model for Robot Learning: A Comprehensive Survey (2026).

Official / primary links: arXiv · arXiv HTML. A recent survey on World Models in robotics, including paradigms, applications, limitations, and links with planning. Use in this dissertation: 2025–2026 state of the art, embodied robotics, benchmarks, and research perspectives. ↩ Back to the bibliography index

A Comprehensive Survey on World Models for Embodied AI (2025).

Official / primary links: arXiv. A survey on World Models for embodied AI. Use in this dissertation: Appendix A.3, robotics, simulation, and embodied AI. ↩ Back to the bibliography index

RAG, Tools, Agents, and Memory

Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.

Official / primary links: arXiv · NeurIPS PDF. A foundational reference for RAG, useful for distinguishing document retrieval from causal understanding. Use in this dissertation: external document memory and the limits of RAG as a substitute for understanding. ↩ Back to the bibliography index

Schick, T. et al. (2023). Toolformer.

Official / primary links: arXiv · ACM Guide. A reference on how language models can learn to use tools. Use in this dissertation: Tool-using Agents, APIs, retrieval, and external computation. ↩ Back to the bibliography index

Yao, S. et al. (2023). ReAct: Synergizing Reasoning and Acting in Language Models.

Official / primary links: arXiv · Google Research Blog · Project / code. A major reference on the integration of reasoning and action in LLM agents. Use in this dissertation: interleaving reasoning and action, Tool-using Agents, and hallucination reduction through interaction. ↩ Back to the bibliography index

Huang, X. et al. (2024). Understanding the Planning of LLM Agents: A Survey.

Official / primary links: arXiv. This survey is useful for planning, memory, reflection, plan selection, and external modules in LLM agents. Use in this dissertation: mapping planning mechanisms in LLM agents. ↩ Back to the bibliography index

Zhang, Z. et al. (2024). A Survey on the Memory Mechanism of Large Language Model based Agents.

Official / primary links: arXiv · ACM TOIS · Associated GitHub. A reference on memory mechanisms in LLM agents. Use in this dissertation: external memory, Agentic Memory, design, and evaluation. ↩ Back to the bibliography index

Du, P. (2026). Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers.

Official / primary links: arXiv · arXiv HTML. A recent survey on autonomous agent memory, including consolidation, recall, forgetting, contradiction, and multimodal memory. Use in this dissertation: Agentic Memory, the write–manage–read loop, and cognitive continuity. ↩ Back to the bibliography index

Neuro-symbolic AI

Garcez, A. d’Avila, & Lamb, L. C. (2023). Neurosymbolic AI: the 3rd wave.

Official / primary links: DOI — Artificial Intelligence Review · Garcez author page. A useful reference for explaining the integration of neural learning and symbolic reasoning. Use in this dissertation: reasoning, rules, explainability, logic, and learning. ↩ Back to the bibliography index

Colelough, B. C., & Regli, W. (2025). Neuro-Symbolic AI in 2024: A Systematic Review.

Official / primary links: arXiv · CEUR Workshop PDF. A recent systematic review of Neuro-symbolic AI. Use in this dissertation: state of the art in Neuro-symbolic AI, gaps, explainability, and metacognition. ↩ Back to the bibliography index

Yang, X.-W. et al. (2025). Neuro-Symbolic Artificial Intelligence: Towards Improving the Reasoning Abilities of Large Language Models.

Official / primary links: arXiv · IJCAI PDF. A survey on the use of Neuro-symbolic AI to strengthen the reasoning capabilities of LLMs. Use in this dissertation: Symbolic→LLM, LLM→Symbolic, and LLM+Symbolic architectures. ↩ Back to the bibliography index

Active Inference

Friston, K. (2010). The Free-Energy Principle: A Unified Brain Theory?

Official / primary links: Nature Reviews Neuroscience · PubMed. A foundational reference on the Free-Energy Principle. Use in this dissertation: perception, action, learning, and uncertainty minimization. ↩ Back to the bibliography index

Friston, K. et al. (2025). Active inference and artificial reasoning.

Official / primary links: arXiv. A recent work connecting Active Inference, reasoning, action selection, and World Models. Use in this dissertation: action selection to reduce uncertainty in World Models. ↩ Back to the bibliography index

de Vries, B. (2026). Active Inference for Physical AI Agents — An Engineering Perspective.

Official / primary links: arXiv. A recent reference on Active Inference applied to physical agents. Use in this dissertation: physical agents, real-time constraints, message passing, and control. ↩ Back to the bibliography index

Predictive Artificial Intelligence Architectures — Causality

Pearl, J. (2009). Causality: Models, Reasoning, and Inference.

Official / primary links: Cambridge University Press · Academic PDF. A foundational reference on causality, interventions, and counterfactual reasoning. Use in this dissertation: distinction between correlation and causality, intervention, and counterfactual reasoning. ↩ Back to the bibliography index

Schölkopf, B. et al. (2021). Toward Causal Representation Learning.

Official / primary links: arXiv · Max Planck publication. An important reference on causality, representations, and out-of-distribution robustness. Use in this dissertation: learning high-level causal variables from low-level observations. ↩ Back to the bibliography index

Predictive Artificial Intelligence Architectures — Model-Free Reinforcement Learning, MCTS, and AlphaZero

Kocsis, L., & Szepesvári, C. (2006). Bandit Based Monte-Carlo Planning.

Official / primary links: Springer (ECML 2006). The seminal publication introducing the UCT algorithm, which established Monte Carlo Tree Search (MCTS) as a practical planning method. Use in this dissertation: search-based planning, decision-making under uncertainty, and predictive exploration. ↩ Back to the bibliography index

Silver, D. et al. (2018). A General Reinforcement Learning Algorithm that Masters Chess, Shogi, and Go through Self-Play.

Official / primary links: Science · Google DeepMind. Introduces AlphaZero, combining deep reinforcement learning with Monte Carlo Tree Search to achieve superhuman performance through self-play. Use in this dissertation: predictive planning, search-guided decision-making, and model-based optimization. ↩ Back to the bibliography index

Predictive Artificial Intelligence Architectures — Cybersecurity, Digital Identity, IoT, and Safety

OWASP Foundation (2025). OWASP Top 10 for Large Language Model Applications 2025.

Official / primary links: OWASP GenAI Security Project. The leading community reference describing the principal security risks affecting LLMs, agentic AI systems, retrieval-augmented generation (RAG), and tool-using agents. Use in this dissertation: AI cybersecurity, prompt injection, excessive agency, model security, supply-chain risks, and Trust Governance. ↩ Back to the bibliography index

National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0).

Official / primary links: NIST AI RMF. The U.S. reference framework for governing AI risks throughout the system lifecycle. Use in this dissertation: AI Governance, Trust Governance, risk management, evaluation, and oversight of Predictive Artificial Intelligence Architectures. ↩ Back to the bibliography index

National Institute of Standards and Technology (2024). Cybersecurity Framework (CSF 2.0).

Official / primary links: NIST Cybersecurity Framework 2.0. The reference framework for cybersecurity governance and organizational resilience. Use in this dissertation: cybersecurity governance, Cyber-Physical Trust, resilience, and Trust Continuity. ↩ Back to the bibliography index

National Institute of Standards and Technology (2020). SP 800-207 — Zero Trust Architecture.

Official / primary links: NIST SP 800-207. Defines the Zero Trust Architecture model based on continuous verification and context-aware access control. Use in this dissertation: Trust Continuity, adaptive trust decisions, Digital Identity, and Cyber-Physical Trust. ↩ Back to the bibliography index

National Institute of Standards and Technology (2025). SP 800-63-4 — Digital Identity Guidelines.

Official / primary links: NIST SP 800-63-4. The latest NIST guidance on identity proofing, authentication, federation, and authenticator assurance. Use in this dissertation: Digital Identity, continuous authentication, Trust Continuity, and identity governance. ↩ Back to the bibliography index

World Wide Web Consortium (W3C). Web Authentication: WebAuthn Level 3.

Official / primary links: W3C WebAuthn Level 3. The web standard for public-key authentication. Use in this dissertation: phishing-resistant authentication, Digital Identity, and comparison with sovereign trust architectures. ↩ Back to the bibliography index

FIDO Alliance. Passkeys.

Official / primary links: FIDO Alliance — Passkeys. Official documentation describing passwordless authentication based on public-key cryptography. Use in this dissertation: passwordless authentication, Digital Identity, phishing resistance, and comparison with EviSKMS. ↩ Back to the bibliography index

European Telecommunications Standards Institute (ETSI). ETSI EN 303 645 — Cyber Security for Consumer Internet of Things.

Official / primary links: ETSI EN 303 645. The European baseline cybersecurity standard for consumer IoT devices. Use in this dissertation: IoT security, Digital Identity, Cyber-Physical Trust, and connected devices. ↩ Back to the bibliography index

National Institute of Standards and Technology (2020). NISTIR 8259A — IoT Device Cybersecurity Capability Core Baseline.

Official / primary links: NISTIR 8259A. Defines the core cybersecurity capabilities expected from IoT devices. Use in this dissertation: device identity, secure lifecycle management, and Cyber-Physical Trust. ↩ Back to the bibliography index

European Union (2024). Regulation (EU) 2024/1689 — Artificial Intelligence Act.

Official / primary links: EUR-Lex — AI Act. The European regulatory framework governing AI systems according to risk categories. Use in this dissertation: AI Governance, Trust Governance, compliance, and high-risk AI systems. ↩ Back to the bibliography index

European Union (2024). Regulation (EU) 2024/2847 — Cyber Resilience Act.

Official / primary links: EUR-Lex — Cyber Resilience Act. Establishes cybersecurity requirements for products with digital elements throughout their lifecycle. Use in this dissertation: cybersecurity by design, IoT, Digital Identity, Cyber-Physical Trust, and lifecycle governance. ↩ Back to the bibliography index

European Union Agency for Cybersecurity (ENISA). ENISA Threat Landscape 2025.

Official / primary links: ENISA Publications. Annual European analysis of emerging cyber threats and evolving attack trends. Use in this dissertation: cyber risk evolution, AI-enabled threats, Cyber-Physical Trust, and operational resilience. ↩ Back to the bibliography index

European Union Agency for Cybersecurity (ENISA) (2025). ENISA Threat Landscape 2025.

Official / primary links: ENISA Threat Landscape 2025 · Official ENISA PDF. The annual European assessment of the cyber threat landscape, documenting emerging attack trends, adversary capabilities, and major incidents. Use in this dissertation: European cyber context, converging threats, and justification of the applied cybersecurity perspective. ↩ Back to the bibliography index

National Institute of Standards and Technology (2020). NISTIR 8259A — IoT Device Cybersecurity Capability Core Baseline.

Official / primary links: NISTIR 8259A · Official NIST PDF. Defines the baseline cybersecurity capabilities expected of Internet of Things devices, including device identity, secure configuration, data protection, logical interfaces, software updates, and cybersecurity state awareness. Use in this dissertation: IoT Digital Identity, device lifecycle, hardware attestation, and secure maintenance. ↩ Back to the bibliography index

European Telecommunications Standards Institute (ETSI) (2024). ETSI EN 303 645 — Cyber Security for Consumer Internet of Things.

Official / primary links: Official ETSI EN 303 645 (Version 3.1.3). The leading European cybersecurity standard for consumer IoT devices, covering default credentials, vulnerability management, software updates, personal data protection, and attack surface reduction. Use in this dissertation: IoT security, connected devices, minimum cybersecurity requirements, and secure lifecycle management. ↩ Back to the bibliography index

FIDO Alliance. Passkeys and FIDO Authentication.

Official / primary links: Passkeys · FIDO Specifications. Official industry references for phishing-resistant passwordless authentication based on public-key cryptography without shared server-side secrets. Use in this dissertation: Digital Identity, human authentication, local proof of possession, and phishing resistance. ↩ Back to the bibliography index

World Wide Web Consortium (W3C) (2026). Web Authentication: An API for Accessing Public Key Credentials — Level 3.

Official / primary links: WebAuthn Level 3 Specification · W3C Candidate Recommendation Announcement. Defines the WebAuthn API enabling web applications to create and use attested public-key credentials bound to a relying party. Use in this dissertation: Passkeys, strong authentication, phishing resistance, and verified Digital Identity. ↩ Back to the bibliography index

European Commission. European Digital Identity Wallet (EUDI Wallet).

Official / primary links: European Digital Identity Wallet · Architecture and Reference Framework. The European framework for user-controlled digital identity wallets supporting selective disclosure and cross-border interoperability under eIDAS 2. Use in this dissertation: sovereign Digital Identity, identity wallets, user consent, and verifiable attributes. ↩ Back to the bibliography index

National Institute of Standards and Technology (2022). SP 800-218 — Secure Software Development Framework (SSDF) Version 1.1.

Official / primary links: NIST SP 800-218. Defines recommended practices for secure software development throughout the software lifecycle. Use in this dissertation: secure development of AI agents, software tools, dependencies, and software supply chains. ↩ Back to the bibliography index

Cybersecurity and Infrastructure Security Agency (CISA). Secure by Design.

Official / primary links: CISA Secure by Design. Promotes shifting cybersecurity responsibility toward software vendors through security-by-design and security-by-default principles. Use in this dissertation: Secure by Design, AI systems, connected devices, and critical software engineering. ↩ Back to the bibliography index

European Union (2024). Regulation (EU) 2024/2847 — Cyber Resilience Act.

Official / primary links: EUR-Lex — Regulation (EU) 2024/2847. Establishes horizontal cybersecurity requirements for products with digital elements throughout their lifecycle. Use in this dissertation: connected devices, digital products, secure lifecycle management, vulnerability management, and European regulatory compliance. ↩ Back to the bibliography index

European Union (2024). Regulation (EU) 2024/1689 — Artificial Intelligence Act.

Official / primary links: EUR-Lex — Regulation (EU) 2024/1689 · EUR-Lex Summary. The European regulatory framework governing AI systems according to risk categories. Use in this dissertation: AI Governance, high-risk AI systems, safety, human oversight, and traceability. ↩ Back to the bibliography index

ISO/IEC 30107. Information Technology — Biometric Presentation Attack Detection.

Official / primary links: ISO/IEC 30107-1:2023. The ISO/IEC family of standards defining terminology and evaluation methods for biometric Presentation Attack Detection (PAD). Use in this dissertation: authentication of living persons, biometrics, liveness detection, PAD, deepfakes, and presentation attacks. ↩ Back to the bibliography index

National Institute of Standards and Technology. Face Recognition Vendor Test (FRVT).

Official / primary links: NIST FRVT. The NIST evaluation program for facial recognition technologies, providing independent performance assessments across diverse operational scenarios. Use in this dissertation: biometric evaluation, human Digital Identity, and limitations of facial recognition systems. ↩ Back to the bibliography index

Predictive Artificial Intelligence Architectures — Appendices

The appendices gather material useful for submission, defense, or external positioning of the dissertation, without overloading the main scientific argument: comparative positioning against the state of the art (Appendix A).

Appendix A — Comparative Positioning Against the State of the Art

Predictive Artificial Intelligence Architectures — A.1. Benchmark Status

This benchmark is not an experimental benchmark of algorithmic performance. Rather, it is a documentary, conceptual, and methodological benchmark intended to position this dissertation in relation to major publications and surveys in the field.

It compares the dissertation with three families of sources:

1. scientific publications specializing in World Models, LLM agents, memory, Neuro-symbolic AI, Active Inference, causality, and reinforcement learning;
2. cybersecurity, Digital Identity, and governance frameworks produced by reference organizations;
3. synthesis documents that map a single subfield without proposing a transversal unifying framework.

The objective is to determine whether the dissertation provides distinct value: not by replacing these works, but by connecting them within a common framework oriented toward Predictive Artificial Intelligence Architectures, memory, causality, planning, cybersecurity, safety, and Trust Continuity.

A.2. Comparison Criteria

The benchmark uses nine criteria.

A.3. Qualitative Comparison with Major Publications

Predictive Artificial Intelligence Architectures — A.4. Differentiation Matrix

Qualitative scoring: 0 = absent, 1 = weak, 2 = present, 3 = central.

This matrix does not claim that the dissertation is superior to specialized publications within their own domains. The high scores assigned to the present dissertation reflect its cross-disciplinary synthesis function (broad coverage), not experimental superiority in every subfield. A robotics survey remains more precise on robotics; NIST remains more normative on Digital Identity; Ha & Schmidhuber remain more foundational on World Models. Rather, the matrix highlights a difference in function: it does not replace specialized surveys; it connects them within a transversal architecture. See also the digital trust comparison in the companion DNA/EviDNA dissertation, which adopts a more cautious reading of interoperability and standardization.

A.5. Distinctive Contribution of the Dissertation

The dissertation is distinguished by eight contributions.

Contribution 1 — Unifying Framework

It shifts the debate from “World Models versus LLMs” to a broader question: which architectures can connect language, perception, memory, causality, prediction, action, and control?

Contribution 2 — Proposed Taxonomy

The proposed taxonomy of Predictive Artificial Intelligence Architectures classifies architectures according to seven dimensions: language, perception, memory, causality, action, prediction, and planning.

Contribution 3 — LAMP-C Architecture

The LAMP-C architecture proposes a synthetic articulation of language, abstraction, memory, prediction, and causality/control.

Contribution 4 — Cyber-Physical Extension

The LAMP-Cyber dimension applies Predictive Artificial Intelligence Architectures to Trust Continuity among humans, machines, AI agents, and connected devices.

Contribution 5 — From Dissertation to Research Program

The dissertation includes falsifiable hypotheses, an AI-TRL maturity framework, benchmarks, and an applied research program.

Contribution 6 — Patented Lineage and Industrialization Evidence

The dissertation connects the Gen1 Cryptographic Genome with the international segmented-key patent (WO/2018/154258) and a non-sensitive evidence appendix derived from EviSKMS-CryptPeer, with public / confidential / IP classification.

Contribution 7 — Cross-Disciplinary Francophone Positioning

Most specialized publications are in English and segmented by domain. This dissertation offers a structured, interactive, research-oriented synthesis in French.

Contribution 8 — Limitations, Falsifiability, and Public Publication

The dissertation includes a limitations and falsifiability section, a companion DNA/EviDNA dissertation, and a short public version, in order to distinguish demonstration, industrialization, applied research, and validation that remains open.