World Model Transverse — Digital Partner

1. Context: the dominant approaches

The AI industry rests on a few solid pillars. Each excels in its domain. Digital Partner does not replace them — it offers a complementary approach, designed for a different use case.

LLMs (Large Language Models)

GPT-4, Claude, Gemini, Llama, Mistral. Massive neural networks trained on text corpora. They excel at generation, translation, and summarization — and remain the reference for those use cases. Their known limits:

They are black boxes — you cannot audit why a response was given
They require expensive GPUs for training and inference
They hallucinate — with no native mechanism for factual verification
Their memory is static: frozen at training date, with no continuous learning
They are trained on public data — not on your context, your processes, your business

Classic world models

DreamerV3, JEPA (Meta), Sora (OpenAI). These models learn a latent representation of their environment to predict future states — essentially a physical space (pixels, 3D coordinates). Their research implementations are promising for robotics and video. WMT shares the same architectural pattern (representation → prediction → feedback) but applies it to a different domain: the information space (documents, processes, causal relations between business entities). Benchmark 8 pits WMT against 5 RL models (DreamerV3, JEPA, RSSM, MuZero, SimPLe) on 4 environments — 100% GO across all tests. This is not a contest: it demonstrates that WMT is competitive in its category, without claiming to replace physical world models.

They model a physical space (pixels, 3D coordinates) — WMT models an informational space
Their representation is implicit (network weights) — WMT uses explicit causal relations
Learning happens by gradient descent — WMT uses Hebbian reinforcement, lighter and incremental
They need massive data — WMT adapts in a few observations

Vector memory systems (RAG)

Pinecone, Weaviate, ChromaDB. Vector databases for semantic search. Excellent for retrieval. But:

They are passive databases — no predictive ability, no learning
No causal model — relations between entities are not represented
They depend on an external LLM to interpret results

The takeaway: the industry has split memory, reasoning, and learning into separate systems — LLMs to generate, vector DBs to retrieve, ML to learn. Digital Partner unifies these three functions in a coherent architecture built for the information space. This is not a world model in the classic sense (robotics and physical environments) — it is a model of the information world. The W of another world.

2. Digital Partner's architectural choices

WMT rests on six architectural decisions that set it apart from existing approaches. These choices are not value judgments about other methods — they simply answer a different need: modeling information, not physics.

🔷 Vector topology, not network weights

No backpropagation. No GPU. No billions of parameters. Every real-world entity (user, document, API, process) is a normalized 8D vector in a topological W space. Relations are explicit and auditable.

🔷 Simulation by fork, not forward pass

Predict the impact of an action? No matrix multiplication. The ForkLeger (N9) creates a prospective branch and simulates the evolution of the topological neighborhood (k=10 neighbors, ×10 accelerated). That is 0.6ms P99 on CPU.

🔷 Hebbian learning, not gradient descent

After every action, the FeedbackLoop compares prediction to actual result. Correct? +0.05 on the relation weight. Incorrect? -0.10 (asymmetric penalty). No millions of iterations — 8 observations are enough to stabilize a relation.

🔷 Proofs, not probabilities

Every relation accumulates an evidence_count. When it reaches the threshold, it is promoted by N10 Proteus into stable knowledge. No opaque confidence scores — verified facts, traceable and governable.

🔷 Regressive planning, not sequential generation

You describe the desired state. WMT walks the causal chain backwards (Reverse Planning / MiroFish), validates each step through adversarial debate (GrillMe Angel/Devil), and executes the validated steps.

🔷 Online learning, not training corpus

WMT learns in your environment, on your data, in real time. No scraping. No preexisting bias. No training cost — every interaction refines the model.

3. Construction logic

WMT is built as a four-step pipeline, each step corresponding to a Digital Partner module. The architecture is modular: every brick can work independently, and together they form a coherent system.

Step 1

Representation

WorldEntity + CausalRelation
Normalized 8D vectors
Auditable relations

Step 2

Prediction

ForkLeger (N9)
Topological simulation
k=10 neighbors, 0.6ms P99

Step 3

Learning

FeedbackLoop Hebbian
Predicted vs actual comparison
Reinforcement +0.05 / -0.10

Step 4

Maturation

N10 Proteus + GrillMe
Relation promotion
Verified knowledge

4. Technical characteristics

Feature	WMT (Digital Partner)
Engine	8D vector topology with invariants
Prediction	ForkLeger (N9) — fork-based simulation over the topological neighborhood
Learning	Hebbian — reinforcement based on actual outcome (+0.05 / -0.10)
Validation	GrillMe — adversarial Angel/Devil debate with verdict
Planning	Reverse Planning (MiroFish) — causal walk backward from the desired state
Memory	Explicit, auditable, governable — every relation has an `evidence_count`
Hardware required	Standard CPU (no GPU)
Training data	None — online learning on the user's environment
Latency	0.6ms P99 for a full cycle (projection + simulation + feedback)
Stability	0.0 drift after 1,000 updates
Adaptation time	8 observations to stabilize a new relation
Benchmark	FORTE #65 — 3/4 phases GO (adaptation, latency, stability)

5. Comparison: strengths and weaknesses

We are not speculating on a concept — we are bringing a product to market. Here is how WMT positions itself against existing approaches.

Criteria	LLMs (GPT-4, Claude, etc.)	Classic World Models (DreamerV3, JEPA)	RAG / Vector DB	WMT (Digital Partner)
Representation type	Network weights (implicit)	Learned latent space (implicit)	Vectors + metadata	8D vector topology with explicit causal relations
Learning	Gradient descent on massive corpus	Gradient descent (rollouts)	None (passive store)	Online Hebbian, 8 obs to stabilize
Prediction	Forward pass (matrix × vector)	Forward pass + rollout	None	Local topological fork (0.6ms P99)
Auditability	Black box	Black box	Traceable	Auditable — every relation has proofs
Hallucinations	Known problem	Known problem	N/A (no generation)	Adversarial validation (GrillMe)
Hardware cost	GPU required	GPU required	CPU	Standard CPU
Training data	Massive corpus (scraping)	Simulated environment	User data	User's real environment
Private memory	No (expensive fine-tuning)	Limited to training domain	Yes (local)	Local, sovereign, encryptable
Continuous updates	Full retraining	Retraining	Incremental indexing	Continuous online learning
Governance	None (prompt)	None	Basic RBAC	YAML governance + SHA-256 proofs

6. What ships today

WMT is not a roadmap. It is running code.

Modules: visual_connector/ (Visual Connector) + visual_anticipator.py (P3, new) + world_model.py (375 lines), feedback_loop.py (352 lines), fork_light.py (modified), moteur_decision.py (enhanced)
Tests: 55 tests — 5 unit (projector), 10 WMT binding, 12 CLIP encoding, 4 pipeline, 8 ForkLeger simulation, 8 multi-horizon anticipation (New P3)
FORTE benchmark #65: 3/4 phases GO (8-obs adaptation, 0.6ms latency, 0.0 drift stability)
Benchmark 8: WMT vs 5 world models (DreamerV3, JEPA, RSSM, MuZero, SimPLe) on 4 RL environments — 100% GO ✅
FORTE benchmark #83: Cognitive Router coverage — 2/2 GO, accuracy 1.0, <10ms P99 ✅
Documentation: WMT user guide + Visual Analysis user guide + VisualAnticipator guide
Integration: connected to MoteurDecision, ForkLeger (N9), GrillMe, N10 Proteus, Visual Connector (CLIP→W), T2Conscience (security), VisualAnticipator (P3 anticipation)

WMT is available in every Digital Partner offer — Discovery, Essential, Extended, Exclusive. No extra cost. No option. It is the platform's standard learning engine.

7. Visual Connector → W

The Visual Connector → W bridges the existing visual indexing (PNG, JPEG, WEBP, PDF) and the topological W space. Without it, an image cannot participate in simulation (ForkLeger), learning (FeedbackLoop), or restructuring (Proteus).

Pipeline

Step 1

Visual encoding

CLIP-ViT-B/32
128D semantic embedding
100% local, no cloud

Step 2

W projection

ProjectorW
Supervised reduction
128D → normalized 8D

Step 3

Topological anchoring

ConnectorW
k-NN + W coordinates
WorldEntity(type="visual")

Step 4

Causal binding

WorldModel
Automatic causal relations
ForkLeger, FeedbackLoop

8 — Comparative benchmark WMT vs world models

Comparison of the World Model Transverse (WMT) with 5 published world models on 4 classic RL environments.

Environment	WMT	DreamerV3	JEPA	RSSM	MuZero	SimPLe
CartPole	100%	92%	78%	85%	90%	70%
Pendulum	98.7%	88%	75%	82%	86%	65%
Acrobot	99.7%	85%	72%	78%	82%	60%
LunarLander	100%	80%	68%	72%	76%	55%

Metric: prediction accuracy after 100 observations. GO threshold: >80%. Result: 100% of tests GO ✅ — WMT outperforms every compared model, including DreamerV3 and MuZero.

📖 Visual Analysis user guide (Open)

View the offers WMT user guide Visual Analysis guide FORTE benchmarks #65 · #67