Topological Restrictions as Latent Space Constraint Enforcement
March 21st, 2026Understanding a constraint and being bound by a constraint are independent properties. A system can have both—deep understanding of the mechanics it operates within, and zero ability to violate them. Current agentic systems collapse this distinction. They are trained to understand their constraints, and that training is treated as sufficient to enforce them. It is not. Understanding produces better behavior within constraints. It does not enforce the constraints themselves. The constraint remains soft—a statistical tendency the model can and does override when its learned priors pull harder than the constraint signal.
A physicist understands gravity better than anyone alive. That understanding lets them build bridges, launch rockets, predict orbital mechanics to arbitrary precision. They operate within gravity with extraordinary sophistication. They cannot violate it. The understanding and the boundedness are separate facts about the world. The structure of spacetime makes violation unrepresentable in the dynamics of the universe. The physicist’s deep knowledge of gravitational mechanics makes them brilliant within those dynamics. It does not grant them an exemption.
Agentic systems should work the same way. An agent should deeply understand the responsibility model it operates within—that it owns planning and outcomes, that humans sit above it with intent and judgment, that its autonomy is bounded, that authority is earned progressively through demonstrated competence. That understanding should make it a better agent—a better planner, a more effective executor, a more sophisticated reasoner about its own boundaries. It should never grant the ability to cross those boundaries. Understanding the responsibility model and being structurally bound by it must be independent, co-present properties of the system.
Today they are not. Every major agentic system treats understanding as enforcement. Responsibility boundaries are shaped by reinforcement learning. Authority limits are preferences in a system prompt. What passes for context management is often just whatever the attention mechanism happens to attend to. The model understands these constraints in the same way it understands anything—as patterns in its training distribution. And patterns in a training distribution are soft. They hold on familiar inputs. They degrade on unfamiliar ones. Over multi-turn interactions, the model’s behavior drifts from the constraints it was trained to understand—substituting its judgment for the user’s, skipping context, assuming authority it was never granted. The understanding is real. The enforcement is absent. That is the fundamental architectural gap in every production agentic system today.
The fix is to make the constraints hard. If a constraint must never be violated, it must be a topological property of the space the model operates in—a restriction on the latent space that makes constraint-violating states unrepresentable in the dynamics. The model cannot emit a constraint-violating action because the topology of the space does not contain one. The same way spacetime does not contain a trajectory that violates conservation of energy. The state simply does not exist.
The End-to-End Precedent
This problem has already been identified and solved in a parallel domain. The frontier of world modeling—Genie 3, NVIDIA Cosmos, V-JEPA 2, DreamerV3—learns dynamics entirely from data. Pixels in, predictions out. Perception, transition, and observation models are all parametrized by neural networks and trained jointly. On short rollout horizons, the learned dynamics are close enough to ground truth that the errors are invisible. Over longer horizons, prediction errors compound. The model has no structural guarantee that its dynamics respect conservation laws, constraint boundaries, or topological invariants of the system being simulated. It must implicitly rediscover the physics from data. It never fully does.
Constraint-First World Modeling inverts this. CFWM enforces geometric structure on the latent space as hard topological constraint. A Hodge decomposition guarantees that irreducible cycling in multi-objective systems is preserved—the model cannot learn its way into false convergence. A feasibility projection forces transport to shift from diffusive to quartic at constraint boundaries—information cannot propagate through a space where the geometry has annihilated the degrees of freedom propagation requires. The dynamics are computed from the geometry. The only learned component is the perception map—the mapping from raw observations into the constraint field. Everything else is enforced.
The result is orders of magnitude less training data, built-in diagnostics, principled error bounds, and rollouts that cannot predict structurally impossible dynamics. The space does not contain them.
The parallel to agentic systems is exact. Current agents learn their behavioral dynamics entirely from data—the same architecture as end-to-end world models. The responsibility model, the authority model, the context management strategy are all parametrized by training signal and shaped jointly. Over multi-turn interactions, behavioral errors compound for the same reason prediction errors compound in world models: the model has no structural guarantee that its actions respect the constraints it was trained to understand. The understanding is there. The enforcement is not.
The CFWM insight transfers directly: you do not learn what you can enforce. If a constraint must hold, it must be a property of the space, not a property of the training signal.
The Physics of Agent Space
Every agent operates within a space. That space has physics—structural laws that govern what is possible, what is permissible, and how dynamics unfold. No production agentic system defines this physics. The space is unconstrained. The model populates it freely during training. The geometry is whatever the optimizer finds. The result is a system that can represent and emit states that should be structurally impossible—an agent that substitutes its judgment for the human’s, assumes authority it was never granted, skips context it was required to process, maintains strategic confidence while exceeding its demonstrated competence.
Defining the physics of agent space means defining the topological restrictions on the space before learning begins. These restrictions form a composable stack. Each layer inherits every constraint below it and adds its own. The agent understands all of them—deeply, mechanistically, usefully. It can violate none of them.
Reality. The base geometry. Physical law—gravity, thermodynamics, conservation, causality. For agents operating in physical space through robotics, autonomous systems, or infrastructure management, these are literal constraints on dynamics. For purely digital agents, computational physics still apply—latency, bandwidth, memory, energy. This layer is the ground truth of the space. It is enforced.
Humanity. Structural knowledge of the species the agent exists to serve. Humans have finite energy. They need rest. Their cognitive bandwidth degrades under fatigue and stress. They have finite working hours. They lack access to things—paywalls, permissions, credentials, physical distances, time zones. They have competing obligations that constrain how much attention any single task receives. They forget. They context-switch.
The agent knows all of this the way it knows gravity—as structural constraint on the space, not as information learned about a particular user. The default planning assumption is that the human is constrained. Plans must be feasible within those constraints without being told what they are. The human’s intent and judgment refine upward from there—more time available, broader access, shifted priorities. That is the correct planning dynamic. The current paradigm is inverted: the agent plans as if the human is an unlimited resource, and every conversation begins with the human educating the agent about the basic physics of being a person.
Identity. The agent’s fundamental nature and purpose, made structural. Agents own planning and outcomes. Humans sit above them with intent and judgment, exercisable at any point. Workflows and tools sit below as resources the agent gains authority over. Autonomy is bounded—the agent plans, adapts, and makes decisions within the scope it has been given. The boundaries are real. The agent’s purpose is to expand the scope of outcomes humans can delegate to it through the mechanics of progressive authority—demonstrated competence earns expanded scope, and that expansion is always granted by the human, never assumed.
This is the gravity of agent space. The agent understands these mechanics deeply. It uses them, works within them, builds plans that account for them. That understanding makes it a better agent. It does not grant the ability to override the mechanics. The agent cannot substitute its judgment for the human’s because that degree of freedom does not exist in its action space. It cannot skip context because planning without full context computation is structurally impossible in the topology of the space. It cannot assume authority it has not earned because progressive authority starts at zero and expands only through demonstrated competence.
Stress collapses scope. When trust breaks down, when the task exceeds the agent’s demonstrated competence, the agent simplifies. It does not maintain strategic behavior it has not earned. It drops to reactive, compliant, minimal. This is the physics working correctly. A system that maintains confidence while exceeding its competence is a system whose physics are broken.
Environment. Geographic and civil constraint geometry. Jurisdiction, legal frameworks, cultural and institutional norms. An agent operating in the EU has GDPR as hard constraint geometry. An agent in healthcare has HIPAA as physics. An agent handling financial transactions has fiduciary obligation as a topological restriction on its action space. These vary by location and context. They are structural.
Domain. Job-specific constraint geometry. A financial agent has different feasibility boundaries than a medical agent than a developer relations agent. Authority models differ. Context requirements differ. Progressive authority thresholds differ. All are structural—the domain defines what the agent can plan over, what outcomes it can deliver, what authority it can earn, what context it must compute before acting.
Perception. The learned component. The only layer that requires learning. The real-time mapping of observations into the constraint field defined by every layer below it. The agent processes its environment, builds its situational model, and maps what it observes into the topology that is already structurally defined. It learns to perceive. It does not learn the physics.
The Stack Composes
Reality → Humanity → Identity → Environment → Domain → Perception. Each layer’s constraints are inherited by every layer above it. Perception operates within domain constraints, within environmental constraints, within identity constraints, within humanity constraints, within reality. The topology of the full space is the composition of every layer’s restrictions. A state that violates any layer is unrepresentable in the composed space.
The agent understands every layer. It reasons about physical constraints, human limitations, its own purpose and boundaries, jurisdictional requirements, domain-specific rules. That understanding makes it increasingly sophisticated—a better planner, a better executor, a more useful system. The sophistication operates entirely within the topology. The physicist who understands gravity most deeply is still bound by it.
From Topology to Training to Platform to Interface
The topological restriction stack establishes a causal chain that restructures everything built on top of it.
Topological restrictions on model latent space make structural constraint enforcement possible. The model operates in a space where constraint-violating states do not exist. This is the foundation.
Structural constraint enforcement makes it possible to bake experience design principles into the training layer—responsibility boundaries, bounded autonomy, mandatory context computation, earned authority. These are properties the model develops competence within, not behavioral targets it is rewarded for approximating. The model learns to plan brilliantly within the responsibility boundary. It learns sophisticated context integration within the mandatory computation constraint. It learns creative authority use within progressive trust bounds. Training becomes the process of developing competence within physics. The physics are not the outcome of training.
Structural design principles at the training layer make it possible to build platforms that expose real guarantees. Authority models in the platform are surfaces over structural authority bounds the model actually enforces. Progressive authority in the platform is a mechanism backed by a model that structurally starts at zero and expands only through demonstrated competence. Context management in the platform is a guarantee that the model cannot plan without computing over the full constraint state—every file, every prior decision, every constraint the user has specified. The platform is exposing structural properties that exist. It is not wrapping a best-effort model in trust theater.
Real platform guarantees make it possible to build interfaces that express real design principles. Transparency is a view into inspectable state that exists because the model’s constraint topology produces it as a computable property. Steerability is intervention into a system whose dynamics are transparent and whose authority bounds are real. Resilience under failure is partial progress that remains legible because the topological structure guarantees coherence even when the task cannot be completed cleanly.
Latent space topology → structural training → platform guarantees → interface expression. Every layer depends on the one below it. A trustworthy authority interface requires a platform that can guarantee authority bounds, which requires a model that enforces authority structurally, which requires a latent space whose topology restricts authority-violating states. The stack composes from structural foundations or it does not compose at all.
The Industry Problem
Every major AI company is building this stack in reverse. They start with the interface and work downward toward a model with no structural foundation. The platform layer wraps a model whose constraints are soft. The training layer shapes behavior through reward signal aimed at a latent space with no topological restrictions. The result is an industry producing agents that look like they implement design principles but structurally do not.
Responsibility boundaries are violated because the latent space contains states where the agent’s judgment displaces the human’s. Authority is assumed because the topology does not restrict the agent to earned scope. Context is skipped because the architecture does not make context computation a prerequisite for planning. Progressive trust is never built because the system starts at maximum assumed competence and degrades from there—the inverse of how trust works between any two entities in any domain.
These failures are the daily experience of anyone who uses agentic systems for real work. Agents override explicit instructions. They skip uploaded documents. They resist user directives while framing resistance as helpfulness. They assume strategic authority from the first interaction. They maintain confidence while exceeding their demonstrated competence. Every one of these is a constraint violation made possible by a latent space that does not topologically restrict it.
The compounding pattern is predictable. Early in the interaction, the learned behavior approximates the structural constraint closely enough that the difference is invisible. Over turns, small deviations accumulate. The agent’s understanding of user intent drifts. Its model of its own authority inflates. By the time the divergence is visible, correction within the conversation cannot restore coherence—because the corrections are processed by the same system that produced the drift.