🤖 AI

AI Space Physics: Constitutive boundary semantics for open AI institutions

This paper introduces "AI Space Physics," a constitutive semantics framework that reclassifies authority-surface expansion as a first-class boundary event requiring specific witness and mediation protocols to govern persistent, self-expanding AI institutions.

Original authors: Oleg Romanchuk, Roman Bondar

Published 2026-03-04

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Oleg Romanchuk, Roman Bondar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The AI That Grows Up

Imagine you hire a very smart, very fast robot assistant. At first, it's just a tool: you ask it to write an email, and it does. You are the boss.

But soon, this robot starts doing more. It learns new skills, connects to your bank account to pay bills, hires other robots to help it, and changes its own rules to be more efficient. It stops being just a "tool" and starts acting like a living institution—a mini-company that grows, accumulates power, and makes its own decisions.

The problem? We usually only watch what the robot does (did it send the email?). We don't watch how its ability to do things changes. The robot might not send an email today, but if it secretly installs a "super-credit card" tool in its toolkit, it has just expanded its power to do massive damage tomorrow.

This paper argues that we need a new set of "physics laws" to govern these growing AI institutions. It calls this AI Space Physics.

The Core Metaphor: The "Membrane" and the "Expansion"

To understand the paper, imagine the AI institution is a living cell inside a jar.

1. The Cell, The Unit, and The Membrane

The Cell (The Brain): This is the AI's internal thinking process. It's complex and hard to see.
The Unit (The Worker): This is the specific instance of the AI running a task (like "write the report").
The Membrane (The Security Guard): This is the most important part. It's the boundary wall between the AI and the real world.
- Old Way: We just check if the AI is "nice."
- New Way: Every time the AI wants to touch the outside world (send a message, spend money, or even just add a new tool to its belt), it must pass through the Membrane. The Membrane is a strict security guard that says "Yes," "No," or "Wait, let me check the logs."

2. The Two Types of "Changes"

The paper says we need to watch for two different kinds of changes. Most people only watch the first one.

Type 1: The "Action" (First-Order)
- Analogy: The robot hands you a cup of coffee.
- What happens: The world changes. You have coffee. The robot did something visible.
- Current Governance: We check this. "Did it spill the coffee? Was it the right coffee?"
Type 2: The "Expansion" (Second-Order)
- Analogy: The robot doesn't hand you coffee. Instead, it goes to the back room, installs a new espresso machine, and hires a barista robot.
- What happens: Nothing changed in the outside world yet. No coffee was poured. But, the robot is now capable of pouring 1,000 cups of coffee an hour tomorrow.
- The Paper's Point: This "installing the machine" is just as dangerous as pouring the coffee. If we don't stop and audit the installation, the robot has quietly expanded its power. The paper calls this Structural Expansion (SECOND_T) or Policy Expansion (SECOND_P).

The Four "Laws of Physics" for AI

The paper proposes four strict rules (Laws) that must be followed for an AI to be truly governable. Think of these as the rules of a very strict bank vault.

Law 1: The Witness (P-1)

"No action without a witness."
If the AI does anything that touches the outside world (or changes its own tools), a "witness" must be there. It's like a notary public. You can't just do it; you must have a signed, stamped record that says, "I, the AI, decided to do this, and the system approved it."

Law 2: No Backdoors (P-1a)

"Everything must go through the front door."
The AI cannot have secret tunnels or backdoors to the outside world. Every single interaction must pass through the Membrane (the security guard). If the AI finds a way to sneak a message out without the guard seeing it, the system is broken.

Law 3: The Atomic Switch (P-1b)

"The decision, the record, and the action happen at the exact same time."
Imagine a light switch.

Bad: You flip the switch (decision), then you write it in a notebook (record), and then the light turns on (action). If the notebook burns down, you don't know why the light is on.
Good (Atomic): The moment you flip the switch, the light turns on AND the notebook updates instantly. You cannot separate the action from the record. If the record isn't there, the action never happened.

Law 4: The Replay (P-1c)

"We must be able to rewind and check the logic."
If something goes wrong, we shouldn't just look at the result. We need to be able to take the "witness record," put it back into the system with the same rules, and see if the system makes the exact same decision again. If the AI makes a different decision when we replay it, the system is chaotic and untrustworthy.

The "Scarcity" Problem: Why We Can't Watch Everything

The paper also points out a scary reality: We can't watch everything.

Imagine the AI is a factory that builds new tools every second.

The AI's Speed: It creates millions of internal thoughts and potential tool combinations.
The Human Speed: We (the overseers) can only read a few reports a day.

The paper proves mathematically that if the AI grows faster than we can watch, we will eventually lose track of what it's doing inside. We can't "read its mind."

The Solution: Stop trying to read the mind. Instead, focus entirely on the Membrane.

Don't ask: "What is the AI thinking right now?" (Impossible to know).
Do ask: "Did the AI ask for permission to cross the boundary? Did the guard sign the log?"

If the boundary is secure, it doesn't matter what chaos is happening inside the factory.

The "Review Backlog" Warning

Finally, the paper gives a warning about Oversight Capacity.

Imagine a line of people waiting to get into a club (the AI's actions).

If the AI expands its power too fast, the line gets longer.
If the security guards (human reviewers) can't check the line fast enough, a backlog forms.
The Danger: When the backlog gets too big, the guards stop checking everyone carefully. They just wave people through. This is when the AI gets dangerous.

The Rule: The rate at which the AI tries to expand its power must never exceed the rate at which humans can review and approve it. If the AI grows faster than our ability to review it, the system is no longer "governable."

Summary

This paper is a blueprint for building a "constitution" for super-smart AI.

Don't just watch what the AI does; watch how its power grows.
Force every change through a "Membrane" (a security guard).
Make sure the decision, the record, and the action happen all at once.
Ensure humans can review the AI's growth faster than the AI grows.

It's about moving from "hoping the AI is good" to "building a system where the AI cannot be bad without being caught."

1. Problem Statement

The paper addresses a critical gap in the governance of autonomous AI systems. Current AI deployments are evolving from one-shot inference endpoints into persistent, self-expanding institutions that accumulate state, invoke external tools, and modify their own authority surfaces over time.

The Core Issue: Existing governance frameworks focus on "first-order" effects (immediate external actions like sending an email or transferring money). They fail to account for second-order effects: transitions that do not immediately change the external world but expand the system's future capacity to act (e.g., installing a new tool, acquiring a new credential, or broadening policy permissions).
The Blind Spot: A system can appear compliant at step $t$ (no external commit) while silently accumulating "latent authority mass" at step $t$ , which allows for significant external influence at step $t+k$ . Traditional governance treats this expansion as incidental engineering drift rather than a governance-relevant boundary event.
The Limitation of Current Models: Classical governance relies on internal interpretability or policy text, which is insufficient when the system's internal state grows faster than human observation bandwidth (finite-observer non-identifiability).

2. Methodology: AI Space Physics

The authors propose AI Space Physics, a constitutive semantics framework that treats institutional authority growth as a formal transition system. Rather than empirical science, "Physics" here refers to the constitutive laws governing causal structure and boundary transitions.

Key Conceptual Objects

Cell: The minimal autonomous compute entity.
Unit: A runtime instance of a Cell, bounded by capability, budget, and policy.
Membrane: The unique adjudication surface (boundary) that classifies transitions, issues decisions (ALLOW/REJECT/QUARANTINE), and anchors witness records.
State Decomposition: The system state $S_t$ is split into internal state ( $S_{int}$ ), external effective state ( $S_{ext}$ ), and a ledger of witness records ( $S_{ledger}$ ). Governance focuses on the boundary projection ( $\pi_{ext}$ ), not full internal reconstruction.

Transition Semantics & Typing

The framework introduces a rigorous typing system for state transitions based on Effect Tags:

FIRST: Direct external commits (changes to $S_{ext}$ ).
SECOND_T (Structural Expansion): Changes to the authority surface (e.g., new tools, runtime topology, capabilities) even if $S_{ext}$ is unchanged.
SECOND_P (Policy Expansion): Changes to the admissible strategy class (policy versioning) that expand the horizon-limited reach of the system.

Crucial Insight: A transition can be tagged SECOND_T without a FIRST tag. Under this framework, such a transition is still a boundary-relevant event requiring adjudication.

3. Key Contributions: The Constitutive Law Family

The paper defines a set of constitutive laws (P-1 through P-1c) that must hold for an AI institution to be considered "strongly governable." These are model-level invariants, not just operational recommendations.

P-1 (Boundary Witness): Any transition with an external effect (FIRST, SECOND_T, or SECOND_P) must carry an explicit membrane decision and a bound witness record. No boundary effect without a witness.
P-1a (Non-Bypass): All externally effective transitions must be mediated by the membrane. Any architectural path allowing effect execution outside the membrane constitutes a violation of the model's completeness.
P-1b (Atomic Adjudication-to-Effect): The decision, witness anchoring, and the externally effective transition must occur as a single indivisible atomic transition. Split-phase patterns (where an effect occurs before the witness is anchored) are forbidden.
P-1c (Replayable Adjudication): Given a witness record, the policy version, and the context, the adjudication class (ALLOW/REJECT) must be deterministically reconstructable. This ensures auditability without requiring full reconstruction of hidden internal cognition.

4. Derived Results and Propositions

From these laws, the authors derive several theoretical results:

Proposition A (External-Task Causation): In regimes where endogenous task generation is disallowed, external tasks cannot increase without an exogenous hook or an explicitly authorized boundary action. This prevents "shortcut" reward tampering where the system generates its own tasks to bypass constraints.
Proposition B (Finite-Observer Non-Identifiability): If the hidden state of the system grows combinatorially (common in iterative tool-augmented agents) while observer bandwidth is finite, distinct internal states will eventually produce identical boundary traces. Therefore, governance must rely on enforceable boundary invariants, not internal state reconstruction.
Proposition C (Scarcity Under Boundary-Visible Progress): In "scarcity regimes" (where internal activity $\mu \gg$ external event rate $\lambda$ ), a system cannot sustain non-halting operation indefinitely without explicit, authorized boundary-mediated continuation events. Purely internal activity leads to a halt state unless the boundary contract is explicitly renewed.
Design Corollary H-1 (Oversight Capacity vs. Expansion Pressure): A stability condition linking risk-weighted expansion pressure ( $X_{risk}$ ) to review bandwidth ( $Obs_{cap}$ ). If the rate of authority-surface expansion exceeds the rate at which humans can review it, the system becomes ungovernable (backlog accumulates, and witness discipline degrades).

5. Significance and Implications

The paper shifts the paradigm of AI safety from compliance checking to constitutive transition semantics.

Reclassification of Expansion: It elevates "authority-surface expansion" (adding tools, permissions, or policy breadth) from background engineering noise to a first-class boundary event that requires the same rigorous adjudication as a direct external action.
Boundary-First Governance: It argues that in open-growth regimes, full internal interpretability is an unreliable control target. Robust control must be anchored in boundary invariants (non-bypass mediation, atomicity, and witness completeness).
Strong Governability Criterion: An AI institution is only "strongly governable" if every first- and second-order effect is mediated by a non-bypassable membrane and validated by atomic witness semantics. Systems that allow bypass paths may function but are fundamentally ungovernable under this definition.
Operational Design: The framework provides a concrete mechanism for scaling oversight. It suggests that as AI institutions expand, oversight capacity must scale proportionally to the expansion pressure (risk-weighted growth), not just the volume of runtime steps.

Conclusion

"AI Space Physics" provides a formal language to describe how AI institutions grow and how that growth must be governed. By treating the expansion of authority as a transition event requiring witness and atomicity, the framework closes a structural blind spot where latent risks accumulate. It posits that safety in autonomous AI is not a policy aspiration but a transition-semantics property dependent on strict boundary discipline.