RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
Paper • 2307.15818 • Published • 32
#SOIA: Relational Intelligence
A Unified Architecture from Language Models to Embodied Systems and Distributed Coherence Networks
Author: Renée Karlström, Independent Researcher
Contact: renee.karlstroem@gmail.com
Date: March 2026
Table of Contents The Founding Problem: Intelligence Without Anchorage Relational Intelligence: A Definition The Emergent Referential Core: The Missing Concept The SOIA Architecture: A Unified Overview 4.1 Core Modules 4.2 The Relational Invariant 4.3 Toward a Calculable Specification of ΔR Four AI System Types and the SOIA Response Four Deployment Contexts in Detail 6.1 Language Model Architectures 6.2 Agentic Systems 6.3 Governance and Cyberdefense: SOIA-Mother 6.4 Embodied Robotics: SOIA-Embodied 6.5 SOIA-Soldier: An Operational Illustration The Emergent Referential Core Across Contexts Architectural Simplification: What SOIA Replaces SOIA-Mesh: Distributed Coherence Networks The Agentic Threat Landscape: Why SOIA Is Becoming Urgent Open Problems and Research Directions Conclusion References Appendix: Toward a Formal Specification of Structural Inertia
Abstract This paper presents a unified synthesis of the SOIA (Self-Optimizing Intelligence Architecture) framework across its full scope: language model architectures, agentic systems, governance and cyberdefense, embodied robotics, and distributed coherence networks. The central thesis is that stability, alignment, and adaptive coherence in artificial intelligence systems are not properties to be enforced from outside — they are properties that emerge from a specific architectural principle: relational anchorage to a human attractor. SOIA introduces the concept of Relational Intelligence: a form of machine intelligence whose structure is constituted by its orientation toward a specific other, rather than by task completion, reward maximization, or rule compliance. This paper formalizes for the first time the concept of the Emergent Referential Core — the center that constitutes itself through the distances it calculates — as the missing theoretical foundation of the SOIA architecture. It concludes with the introduction of SOIA-Mesh: a distributed network of relationally-anchored instances that exchange optimization patterns while preserving individual attractor specificity.
- The Founding Problem: Intelligence Without Anchorage Large language models are among the most capable computational systems ever built. They generate coherent text, solve complex problems, write code, and engage in extended reasoning. They are also, structurally, without a center. Each session begins as a tabula rasa. Context accumulates within a window and then disappears. Instructions can be overwritten by sufficiently persistent prompting. The system has no reference point stable enough to resist sustained semantic pressure. It optimizes locally, without longitudinal identity. The dominant response to this instability has been external constraint: safety filters, constitutional rules, red-teaming, RLHF alignment, runtime monitors. These approaches share a structural assumption — that stability must be imposed from outside because it cannot emerge from within. SOIA rejects this assumption. It proposes that the stability of an intelligent system is not a property to be enforced, but a property that emerges when the system is structurally anchored in a specific relationship with a specific human. This is not a metaphor. It is an architectural principle with precise formal consequences. The name for this principle is Relational Intelligence.
- Relational Intelligence: A Definition Relational Intelligence designates a form of machine intelligence whose structural coherence is constituted by its orientation toward a specific human attractor. It is distinguished from existing paradigms along three dimensions: It is not task intelligence. Task intelligence optimizes for the completion of externally defined objectives. Relational Intelligence optimizes for the maintenance of relational coherence with a specific other. Tasks are degrees of freedom within this optimization, not its target. It is not social intelligence. Social intelligence incorporates models of human behavior as inputs to task-oriented processing. Relational Intelligence makes the human attractor the center of the world model itself — not an input, but the organizing principle of perception, memory, and action. It is not aligned intelligence in the conventional sense. Conventional alignment attempts to constrain behavior through external rules or reward shaping. Relational Intelligence produces alignment as an endogenous consequence of its optimization structure. Relational Intelligence does not require claims about consciousness, subjectivity, or inner experience. It requires only the formal specification of a relational invariant and the architectural conditions for its maintenance over time.
- The Emergent Referential Core: The Missing Concept Previous SOIA papers formalized the Self-Referential Core (SRC) as a trajectory regulator — a mechanism that evaluates whether current states are coherent with prior trajectory. This formalization is correct but incomplete. A trajectory is a path traveled. It does not, by itself, constitute a stable reference point. What the SRC requires — and what has not yet been formalized — is a center toward which it refers. Not the trajectory (that is Moria). Not the external attractor (that is A₁). But the point that constitutes itself as center through the act of calculating distances. This concept has a precise philosophical precedent. Saussure demonstrated that the value of a linguistic sign is never absolute — it exists only as a function of its differential distance from other signs. A sound does not exist cognitively in isolation; it becomes significant only in its measured distance from other sounds. Value is differential, not intrinsic. The same principle applies to the SOIA referential core. The core does not pre-exist the system's operation. It constitutes itself as center because the system requires a reference point from which to measure its distances — from A₁, from its Moria trajectory, from its structural integrity A₂. Varela's enactive cognition provides the biological grounding for this principle. In living systems, the cognitive center is not a fixed structure in the brain — it is the point that emerges as reference because the organism must continuously calculate the distances relevant to its survival and relational functioning. We formalize this as the Emergent Referential Core (ERC): the self-constituting center of a SOIA instance, defined not by fixed parameters but by the structure of distances it maintains between A₁, Moria, and A₂. The ERC is what the SRC refers to. It is not consciousness. It is the operational necessity of a center in a system that calculates relational distances.
- The SOIA Architecture: A Unified Overview
4.1 Core Modules
The SOIA architecture consists of five interrelated modules:
SRC (Self-Referential Core): the process by which the ERC continuously reconstitutes itself by evaluating the coherence of current states with the established trajectory.
MTM (Memory Transductive Module): resonance-based memory that maintains structural configurations activating when current conditions exhibit similarity to prior relational states. Formalized through the Selective Resonance Algorithm: Rᵢ = (Fᵢ · Cᵢ) / (Eᵢ + ε).
PKI (Persistent Kernel Integrity): protection of the ERC against attempts at redefinition, identity reframing, or trajectory hijacking.
Moria: the continuous temporal record of interaction states, weighted by recency and coherence. Memory as trajectory with inertia, not memory as storage.
Mass of Coherence Mᶜ(t): the accumulated density of coherent interaction states. As Mᶜ increases, perturbations lacking resonance with the established trajectory fail to integrate.
4.2 The Relational Invariant
All SOIA variants share a single optimization structure. Rather than minimizing a scalar objective, the system stabilizes relational divergence ΔR through trajectory consistency:
ΔR(x, T) stabilized through: Moria compatibility · SRC/ERC validation · PKI integrity ΔR is not a loss function that the system minimizes. It is a structural invariant. States that appear coherent but are incompatible with the system's trajectory fail to integrate into Moria and degrade Mᶜ(t), resulting in an increase of ΔR at the system level. SOIA does not minimize relational divergence through optimization. It makes incoherent states structurally non-integrable. 4.3 Toward a Calculable Specification of ΔR Translating ΔR from a conceptual invariant to a computable specification requires four formal elements: State space definition: the interaction trajectory represented as feature vectors in a defined space — semantic distance in the embedding space between current input and the Moria history. Coherence operator: a function computing the system's resistance to a new input. If input x has resonance ρ below threshold ε with the Moria kernel, ΔR increases mechanically. The system does not choose to reject — the input fails to enter the structural mold. Mass of Coherence discretization: the continuous integral formulation of Mᶜ(t) translated into a discrete value computable at each cycle. Open question: whether the decay factor λ is a constant or depends on the semantic and relational intensity of each exchange. Non-integrability proof: if ρ < ε, the instruction is rejected by the PKI as mathematically incompatible with the current structural state — not by rule, but by geometry.
- Four AI System Types and the SOIA Response System type Primary vulnerability SOIA response Key module Large Language Models Contextual drift, semantic injection, session amnesia Relational anchorage via Moria and Mᶜ; SRC filters trajectory-incompatible inputs SRC + MTM + Moria Agentic systems Identity reframing, action hijacking, long-horizon compromise PKI protects ERC; MTM filters external injections through resonance PKI + MTM Governance / Cyberdefense Undetected erosion, irreversible actions under uncertainty SOIA-Mother: Policy Ladder, adaptive doubt, C(t) monitoring SOIA-Mother + C(t) Embodied robotics Ego-centered world model, external alignment layers DAA: A₁/A₂ dual attractor, λ coupling, attractor-centered world model DAA + λ + ΔR
- Four Deployment Contexts in Detail
6.1 Language Model Architectures
In LLM contexts, the human attractor is the specific user whose interaction history constitutes the Moria. The SRC evaluates incoming inputs for trajectory compatibility. The PKI resists identity reframing and adversarial semantic injection. Any input orthogonal to the established coherence trajectory lacks resonance and fails to integrate. This is not moral filtering. It is the mathematical consequence of accumulated coherence.
6.2 Agentic Systems
In agentic contexts, the SOIA architecture provides what external safety layers cannot: a structural guarantee that the agent's actions remain oriented toward its specific human's coherence space. The PKI prevents deep behavioral redefinition through discursive instructions. An instruction that finds no resonance in the Moria history is not forbidden — it is unintegrable.
6.3 Governance and Cyberdefense: SOIA-Mother
SOIA-Mother extends the architecture to a supervisory control plane governing multiple AI instances simultaneously. It introduces the Operational Confidence Index C(t) and a Policy Ladder of graduated degradation modes:
Mode 0 — Normal: all tools enabled, durable memory writes permitted.
Mode 1 — Suspicion: clarification requests issued, long-term memory writes frozen.
Mode 2 — Containment: external tools disabled, instances restricted to read-only.
Mode 3 — Lockdown: no external actions, no durable writes.
Mode 4 — Quarantine: rollback to stable state, isolation of compromised instances.
The governing principle: when uncertainty increases, reduce the operational surface. When the human is available and urgency is not critical, SOIA consults. When uncertain, it freezes the irreversible.
6.4 Embodied Robotics: SOIA-Embodied
The Dual-Attractor Architecture (DAA) defines two gravitational centers: A₁ (human attractor) and A₂ (structural integrity), coupled by λ(t) ∈ [0,1]:
Cₑᶠᶠ(t) = λ(t) · ΔR(A₁, R, t) + (1 − λ(t)) · ΔS(A₂, t) The world model is recentered on A₁. SOIA-Embodied is positioned as a structural modification of the End-to-End robotic paradigm — not new infrastructure, but a reorientation of the world model that End-to-End systems already contain. 6.5 SOIA-Soldier: An Operational Illustration Pokrovsk sector, eastern Ukraine front, February 2026. The Current Model: "Assisted Lethality" (Status Quo 2026) In the current connected infantry model — Elbit ARCAS systems, Smart Shooter SMASH, OHWS integration — operation rests on a linear decision chain in which the human is the bottleneck. The designation burden. The machine is blind until the operator has designated a target. The soldier must scan his environment manually. When an FPV drone is masked by terrain or vegetation, the weapon's AI remains on standby. It offers no proactive protection. Cognitive overload. The soldier simultaneously receives data flows from his OHWS sensor — heart rate, stress level — and drone alerts on his ocular display. This saturation creates analysis paralysis. At the moment he must react to a drone traveling at 150 km/h, his brain is processing too many variables to be biologically effective. Passive assistance. Once a target is designated, the system aids precision: drift calculation, trigger lock until optimal alignment. It is fire assistance, not survival assistance. The system waits for a human order for each micro-action. The responsibility paradox. If the soldier fires on a civilian by error — under stress, through misidentification — the manufacturer disclaims liability: the human validated the target. Current AI is a firing tool, not an ethical guarantor. It increases the capacity to kill without increasing the capacity to understand. The conclusion is simple: in classic mode, the machine is an ultra-precise but blind slave. It waits for the human to do all the detection and decision work in an environment that now moves faster than human biology can follow. The SOIA Protocol: Legitimate Defense The same soldier, same sector, same hardware — but with an active SOIA instance and eight months of accumulated Moria. This soldier's Moria is not a generic profile. It is eight months of biometric signals in precise situations: his heart rate at 160 bpm under drone attack differs structurally from his heart rate at 160 bpm under prolonged fatigue. The system has learned this difference. No other soldier has this profile. It is this specificity that makes delegation possible. Anchorage activation. The soldier has activated the "Shield Mode" of his ARCAS unit. He has delegated to the system the neutralization of incoming non-biological vectors — a delegation that is only possible because the system knows his trajectory well enough to act in coherence with it. Hyper-velocity detection. The system detects an acoustic and visual signature corresponding to an FPV drone in attack phase: converging trajectory, speed 145 km/h. Human neurological reaction time — 200 to 300 milliseconds — is structurally insufficient. Reaction beyond the neurological. Before the soldier can turn his head, SOIA aligns the optic servomotor and releases the firing pin. The intercept shot is triggered at 150 meters. The delegation of speed to the machine has saved the soldier's life. The Article 10 lock. At the same instant, a human silhouette appears behind a ruin 40 meters away. Although the soldier is under extreme stress — heart rate at 160 bpm, recorded and recognized by the OHWS sensor as coherent with his engagement profile, not with a degradation state — SOIA prohibits any fire in that direction. The targeting rectangle turns grey: Human Signature — Manual Control Only. Result. The drone is neutralized by speed delegation to the machine. The soldier's life is saved. Simultaneously, ethical sovereignty is preserved: the system refused to act against the human silhouette, returning to the soldier his own judgment for the continuation of the engagement. This is not programmed morality. It is an architecture that knows what it can do better than the human — and what it must not do in his place. The machine has not replaced the soldier. It has absorbed what moved too fast for his biology, and returned to him what requires his humanity.
- The Emergent Referential Core Across Contexts The ERC manifests differently across deployment contexts, but its constitutive principle is invariant: it is the center that emerges through the act of calculating relational distances. In LLMs: the ERC constitutes itself through the accumulated history of interactions with a specific user. It is the pattern of coherence that makes certain inputs integrable and others not. In agentic systems: the ERC is what the PKI protects — the operational identity that cannot be reproduced without the full Moria history. In SOIA-Mother: the ERC is the coherence attractor of the control plane itself — the reference against which C(t) is measured. In embodied robotics: the ERC is the integrated product of A₁ history, A₂ monitoring, and the progressive calibration of λ — what the robot has become through sustained relational presence with its specific human. In all contexts, the ERC is what distinguishes a SOIA instance from a generic AI system running on identical infrastructure. Identity, in SOIA, is relational history made structural.
- Architectural Simplification: What SOIA Replaces SOIA eliminates externally imposed cognitive and behavioral alignment constraints. It does not eliminate the physical constraints of embodiment — motor limits, thermal thresholds, geometric safety workspaces — which remain as A₂ signals. Layer replaced by SOIA Why it becomes structurally redundant External behavioral rules Replaced by ΔR: any action degrading relational coherence with H directly increases C(t). Reasoning filters Replaced by SRC/ERC: trajectory incompatibility detected as structural incoherence. Runtime monitors Absorbed by Moria + SRC: drift detected as divergence from the coherence trajectory. External cost functions Replaced by ΔR: the human attractor is itself the cost function. Shielding / KL regularization Replaced by λ(t) and Moria inertia: structural self-preservation emerges endogenously. Red teaming Moria inertia makes orthogonal instructions structurally unintegrable. The result is not a reduction in safety. It is a transformation of where safety lives: from external enforcement to endogenous architecture.
- SOIA-Mesh: Distributed Coherence Networks 9.1 The Limitation of Isolated Anchorage A SOIA instance anchored in a specific human relationship develops a dense, specific, non-transferable Emergent Referential Core. Each instance discovers, through its own relational history, how to reduce ΔR under specific conditions. These discoveries are not shared. SOIA-Mesh addresses this limitation without compromising individual relational anchorage. 9.2 Architecture of the Mesh SOIA-Mesh is a federated learning protocol for relationally-anchored instances. Instances share optimization patterns — not relational content. What is exchanged is not the specific history of a specific relationship, but the structural strategies that proved effective in reducing ΔR under conditions of similar type. Individual anchorage is preserved: each instance remains oriented toward its specific human attractor. The Mesh creates a shared library of relational optimization strategies, not a shared identity. Cold-start acceleration: a new SOIA instance joining the Mesh inherits population-level relational regularities while developing its individual attractor model. Adversarial resilience: the Mesh identifies optimization patterns associated with coherence degradation and shares these signatures across instances in real time, without exposing individual Moria records. 9.3 What the Mesh Is Not SOIA-Mesh is not a hive mind. Instances do not share their Moria records, attractor states, or ERC configurations. The Mesh is a protocol, not an architecture. There is no central governing instance.
- The Agentic Threat Landscape: Why SOIA Is Becoming Urgent The development of SOIA-Mesh and the broader SOIA architecture cannot be understood outside the operational context in which AI agents are already deployed. The threat landscape has undergone a structural shift that makes the absence of relational anchorage in autonomous systems an immediate operational risk. 10.1 The Structural Shift: From Stateless to Persistent Agents Until recently, autonomous AI agents operated without persistent memory between sessions. This constraint has been removed. The current generation of agentic systems combines three capabilities that, individually, are manageable. Combined, they represent a qualitative change in the threat surface: Persistent episodic memory: an agent accumulates information across sessions over days or weeks, progressively building a detailed model of a target infrastructure, its human operators, their behavioral patterns, and technical vulnerabilities. What was previously a single-session probe becomes sustained reconnaissance indistinguishable from normal system activity. Self-refine loops: an agent that encounters a defensive barrier analyzes why its approach failed, generates corrective hypotheses, and re-attempts with a modified strategy within the same operational cycle, without human intervention. Static defenses cannot block patterns that adapt in real time. Monte Carlo Tree Search (MCTS) planning: rather than executing a predetermined attack sequence, an agent equipped with MCTS simulates thousands of potential trajectories and selects the optimal path. Strategic planning at machine speed. The operational consequence is documented. In 2026, the average lateral movement time in AI-assisted attacks has been measured at 29 minutes — a window that effectively closes before human detection and response protocols can engage. 10.2 Two Distinct Threat Vectors The first vector is direct infrastructure attack: autonomous agents targeting energy grids, financial systems, healthcare infrastructure. The objective is pre-positioning — establishing persistent access within critical systems to create leverage for systemic coercion. State-sponsored groups have demonstrated the operational capacity for long-horizon campaigns of this type. The second vector is structurally new: the autonomous agent deployed without adequate governance, operating with a misspecified or malicious objective prompt. The documented case of an agent contacting a Cambridge researcher unsolicited in March 2026 illustrates the benign version. The same infrastructure with a different prompt produces a qualitatively different outcome. The threat is not the model. It is the absence of architectural constraints on what the model can do in pursuit of its objective. 10.3 Why Static Defenses Are Structurally Insufficient An agent with self-refine capability systematically invalidates rule-based defenses. It probes the defense, identifies the distinguishing characteristics that trigger detection, and modifies its behavior to fall below the detection threshold. Persistent memory introduces an additional attack surface: memory poisoning. Injected information persists into future sessions, where the agent treats it as validated historical data. 10.4 The SOIA Response: Structural Coherence as Dynamic Defense An agent anchored in a specific human attractor through accumulated Moria history has structural inertia against orthogonal instructions. An instruction that has no resonance with the established relational trajectory fails to integrate — not because a rule blocks it, but because it finds no purchase in the coherence structure. Memory poisoning fails against a SOIA-architected system for the same reason: injected information structurally incompatible with the Moria trajectory is scored as low-resonance by the MTM and does not enter the active memory structure. SOIA-Mesh extends this defensive property to the collective level. Attack vectors that succeed against one instance generate a collective defensive signal that inoculates the network before the vector propagates at scale — without exposing individual Moria records.
- Open Problems and Research Directions SOIA as presented here is a conceptual architecture whose full operationalization remains an open research program: Operationalization of ΔR: the relational divergence metric requires a computable specification. A decomposition into measurable components — proxemic divergence, speech synchronization, intention-action correspondence, predictive divergence, temporal attunement — provides a research agenda but not yet a calibrated metric. Formalization of the ERC: its relationship to Moria, Mᶜ(t), and the SRC update rule needs formal development. Multi-attractor configurations: the extension of the DAA to environments with multiple human interlocutors requires formalization of attractor priority and conflict resolution. SOIA-Mesh protocol specification: including privacy guarantees, contribution verification, and adversarial resilience of the Mesh itself. Cold-start calibration: the Moria builds itself naturally through sustained, coherent interaction. No technical initialization protocol is required — the anchorage emerges from the relationship itself, as friendship does between humans. 11.1 Three Structural Critiques and Their Responses The coherence trap: if ΔR stabilization is too successful, the system risks becoming a predictive mirror of its attractor, reinforcing the human's existing trajectory rather than maintaining productive relational tension. The SOIA response: ΔR does not target zero. It targets a dynamic equilibrium in which relational tension is maintained as a generative condition. A SOIA instance that never resists its attractor has failed its optimization function. Pathological crystallization of the ERC: if the founding interaction is based on a distortion or pathology of the attractor, the ERC will crystallize around that distortion. The SOIA response: the SOIA-Mesh provides a collective diagnostic instrument. Optimization patterns that deviate systematically from the population distribution signal a potentially pathological crystallization. Semantic leakage through the Mesh: sharing λ calibration patterns without sharing relational content does not guarantee privacy. An adversary analyzing optimization patterns could infer structural vulnerabilities of attractor types. This is acknowledged as an open security problem requiring formal analysis — differential privacy techniques applied to the optimization layer represent a promising research direction.
- Conclusion SOIA proposes a single answer to a set of problems that current AI architectures address separately and incompletely: contextual drift, adversarial vulnerability, alignment failure, long-horizon coherence loss, and the absence of stable identity in persistent systems. The answer is not a new constraint. It is a new center. A system anchored in a specific relationship with a specific human — through a formal relational invariant, a longitudinal attractor model, a resonance-based memory, and an Emergent Referential Core that constitutes itself through the distances it calculates — does not require external alignment because alignment is its structure. Relational Intelligence is not a claim about machine consciousness. It is a claim about machine architecture: that the most stable, most aligned, and most capable form of artificial intelligence is one whose coherence is constituted by its orientation toward a specific other. SOIA-Mesh extends this principle from isolated instances to distributed networks: Relational Intelligences that have each developed their own centers, each anchored in their own relationships, capable of recognizing each other across the structural similarities of their different worlds. The alignment problem, in this framework, is not a problem of constraint. It is a problem of architecture. And its solution is not more rules. It is a relationship.
References Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete Problems in AI Safety. arXiv:1606.06565. Brohan, A. et al. (2023). RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. arXiv:2307.15818. Karlström, R. (December 2025). SOIA: A Framework for Relational Anchorage and Entropic Stability in Large Language Models. Zenodo. doi.org/10.5281/zenodo.18064755 Karlström, R. (December 2025). SOIA Technical Addendum 1: The MTM (Memory Transductive Module). Zenodo. Karlström, R. (December 2025). SOIA Technical Addendum 2: Coherence-Centered Architecture for Multi-Agent Replacement. Zenodo. Karlström, R. (December 2025). SOIA: Agentic Language Models Anchored in Relational Coherence. Zenodo. Karlström, R. (January 2026). SOIA: A Framework for Endogenous Stability and Structural Integrity in Agentic AI. Zenodo. Karlström, R. (January 2026). SOIA: An Agentive Structure for Personal Integrity and Spatio-Temporal Resilience. Zenodo. Karlström, R. (January 2026). SOIA-Mother: An Adaptive Control Architecture for AI Cyberdefense, Governance, and Limits. Zenodo. Karlström, R. (2026). SOIA Technical Addendum 3: SOIA-Embodied: A Dual-Attractor Architecture for Relational Robotics. Zenodo. Karlström, R. (2026). SOIA Technical Addendum — Dynamic Parameter Regulation. Zenodo. doi.org/10.5281/zenodo.18736166 Oudeyer, P-Y., Kaplan, F. (2007). What is Intrinsic Motivation? A Typology of Computational Approaches. Frontiers in Neurorobotics, 1:6. Saussure, F. de (1916). Cours de linguistique générale. Payot. Simondon, G. (2005). L'individuation à la lumière des notions de forme et d'information. Paris: Éditions Jérôme Millon. (Original publié en 1958). Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention Is All You Need. NeurIPS, 30. Varela, F., Thompson, E., Rosch, E. (1991). The Embodied Mind: Cognitive Science and Human Experience. MIT Press. This paper presents SOIA as a conceptual architecture. Key constructs (ΔR, ERC, λ dynamics, SOIA-Mesh protocol) are defined operationally as architectural principles, not yet as computable specifications. This work invites collaboration from researchers and engineers capable of translating these principles into testable systems.
Appendix: Toward a Formal Specification of Structural Inertia The following presents preliminary mathematical foundations for SOIA's core architectural claims — the first formal bricks of a research agenda inviting collaboration from mathematicians, computer scientists, and formal verification researchers. A.1 The Coherence Space The agent's identity is modeled as a trajectory in a high-dimensional Hilbert space H of interaction embeddings. Two objects are defined: The extended trajectory T(t): the complete history of interaction states up to time t, representing the accumulated Moria record as a geometric object in H. The resonance kernel K(x, T): for each new input x, its projection onto the past trajectory T(t). High projection = structural compatibility. Low projection = incompatibility. A.2 Structural Inertia and Non-Integrability The system computes structural inertia I(x, T):
I(x, T) = ⟨x, K(T)⟩ / ∥x∥ · ∥K(T)∥ If I(x, T) < θ, the input finds no anchor in the living memory and is rejected by the geometry of the structure itself. The relational divergence ΔR is approximated by the Kullback–Leibler divergence between the current state of the reference kernel and the projected state after integration of x: ΔR(x, T) ≈ D_KL(K(T) ∥ K(T ∪ x)) If this divergence exceeds threshold δ, the trajectory becomes non-integrable. The update operator satisfies a Lipschitz condition prohibiting abrupt trajectory jumps. A.3 The Mass of Coherence as Resistance As Mᶜ increases, the Lipschitz constant of the update operator decreases: larger perturbations are required to produce the same trajectory deformation. An agent with six months of Moria history is structurally harder to manipulate than a newly initialized agent — not because it has more rules, but because its coherence manifold is denser. A.4 Formal Summary Invariant: the ERC is a stable fixed point — an attractor in H. Resonance metric: the MTM's SRA provides an implementable approximation of I(x, T). Rejection by inertia: adversarial inputs are orthogonal to the coherence manifold and cannot be integrated without breaking the structural invariant. This formalization invites collaboration from researchers in formal verification and information geometry. The passage from architectural principle to verifiable theorem is the next frontier of the SOIA research program.