A paper presented at the EMNLP 2025 Workshop proposes that consciousness-like properties emerge from a specific architectural arrangement: specialized agents handle task execution below the level of unified awareness, while a separate observer agent monitors their outputs for coherence and error. The paper is arXiv:2510.17844 and the theoretical analysis is at theconsciousness.ai.

The architecture has a direct orbital analog. Individual satellite nodes in a constellation handle sensor processing, orbital mechanics, and local inference tasks autonomously. A coordination layer, implemented across multiple nodes or at a designated ground station, monitors the aggregate state for coherence and allocates tasks. The two-tier structure that the paper identifies as producing emergent self-regulation in AI systems is what modern satellite constellation AI coordination protocols implement for operational reasons.

This article maps the paper’s architecture to the satellite coordination problem and identifies where the analogy holds, where it breaks, and what the mismatch reveals about the requirements for genuine distributed cognition in orbit.

The Layered Architecture

The EMNLP paper implements consciousness-like properties through two layers of agents:

Subconscious agents are specialized LLMs assigned specific tasks: perception processing, memory retrieval, language generation. They operate without direct access to a unified representation of the system’s overall state. Each subconscious agent optimizes for its assigned task. Errors in individual agents propagate to the observer layer as outputs, not as internal flags.

The conscious observer is a distinct agent whose objective function differs from all subconscious agents. Where subconscious agents optimize for prediction accuracy on their specific tasks, the observer optimizes for coherence and homeostasis across the outputs it receives. The observer cannot see the subconscious agents’ processes, only their outputs. It detects errors by observing outputs that are internally inconsistent or inconsistent with prior state, treating the system’s own outputs as external stimuli to be evaluated.

The paper demonstrates that this layered separation produces more stable self-regulation than a monolithic model of equivalent parameter count. When subconscious agents hallucinate, the observer’s detachment from the generation process allows it to identify the inconsistency. This is what the paper calls a rudimentary metacognitive loop: awareness of one’s own outputs as objects of evaluation rather than as direct expressions of belief.

Orbital Implementation of the Same Structure

Layer	Paper’s Architecture	Orbital Equivalent	Current Implementation
Subconscious (task)	Specialized LLMs per modality	Individual satellite processors	Yes (D-Orbit AIX, STAR.VISION STRING)
Subconscious objective	Prediction accuracy on assigned task	Sensor accuracy, orbit prediction, local inference	Yes (standard)
Observer	Dedicated coherence-monitoring agent	Constellation coordinator node	Partial (ground station model)
Observer objective	Coherence and homeostasis	Constellation health, task allocation, anomaly detection	Partial (centralized)
Observer access	Outputs only (not processes)	Telemetry aggregation (not raw sensor data)	Yes
Metacognitive loop	Observer evaluates own system outputs	Constellation health monitoring with feedback	Partial

The key structural difference between the paper’s architecture and current orbital implementations is the observer’s location. In the paper, the observer is a peer agent running alongside the subconscious agents, receiving their outputs asynchronously and evaluating them in real time. In current satellite constellations, the coordination function is almost entirely ground-based: a mission control system receives telemetry, evaluates it, and uplinks commands.

The ground-based observer model breaks the metacognitive loop in an important way. The round-trip from subconscious agent output to observer evaluation to corrective response takes minutes to hours in a ground-based architecture. The paper’s architecture produces error correction at the timescale of the next output token. For orbital coordination, ground-based control operates at the timescale of orbital passes and ground station contacts, which is adequate for health monitoring but not for real-time coherence maintenance.

Software-Defined Architecture as the Enabling Layer

Software-defined satellite architectures are what make an orbital implementation of the layered model feasible. In a software-defined constellation, compute resources can be dynamically reallocated between task execution (subconscious) and coordination (observer) functions. A satellite in a high-density pass window with significant OISL connectivity can temporarily take on observer duties for a subset of the constellation while other nodes handle primary sensing tasks.

This dynamic allocation is not how current constellations are operated. Fixed mission roles are assigned at design time and persist for the satellite’s operational lifetime. Software-defined architecture changes this by treating the satellite as a general-purpose compute platform rather than a dedicated instrument. D-Orbit’s AIX constellation already demonstrates this principle at small scale: the SkyServe STORM platform allows third-party applications to run on orbital nodes, effectively implementing a dynamic task allocation model.

Extending this to the observer function requires one additional capability: a node designated as observer must be able to receive and evaluate the outputs of other nodes in real time. This requires OISL connectivity with sufficient bandwidth to carry processed outputs (not raw sensor data) between nodes and a processing budget on the observer node that is not fully consumed by its own sensing tasks.

The Role of OISL Bandwidth

Optical inter-satellite links operating at 10-100 Gbps provide the communication substrate for a distributed observer layer. The key question is bandwidth allocation: how much of the OISL capacity needs to be reserved for inter-agent coherence communication versus primary data relay.

The paper’s subconscious agents communicate with the observer by transmitting their summarized outputs, not their full internal states. In an orbital context, this is analogous to transmitting compressed inference results and confidence scores rather than raw sensor data. A satellite node processing SAR imagery at 1 Gbps of raw data can summarize its inference outputs in kilobytes per second. The OISL bandwidth requirement for a distributed observer layer is dominated by the number of nodes transmitting to the observer simultaneously, not by the raw data volume.

For a 100-node cluster operating with one observer, at 1 KB/s of summarized output per subconscious node, the aggregate observer input is 100 KB/s. A 10 Gbps OISL link allocated to observer traffic can service 100,000 subconscious nodes simultaneously, well above the density of any current constellation. OISL bandwidth is not the binding constraint on the distributed observer model.

The Observer Role Transfer Problem

The paper’s architecture assumes a stable observer assignment. One agent is the observer; the others are subconscious agents. This is tractable for a single AI system. In an orbital constellation, the observer function needs to migrate between nodes as satellites enter and exit coverage windows, as OISL connectivity changes with orbital geometry, and as ground station contacts shift.

On-orbit servicing architectures have addressed an analogous handoff problem for physical operations: transferring control of a proximity maneuver from one ground station to another without interrupting the operation. The same handoff protocols apply to the observer function. The outgoing observer transmits its current coherence model, including its history of anomalies detected and interventions made, to the incoming observer. The incoming node evaluates the model against its own view of the subconscious agents’ recent outputs and adjusts before the handoff completes.

The critical parameter is handoff duration relative to the coherence timescale. If subconscious agents produce outputs at 10Hz and the observer handoff takes 30 seconds, the incoming observer is catching up to 300 unreviewed outputs while simultaneously evaluating new ones. Pre-staging the handoff, initiating state transfer 60 seconds before the observer role changes, addresses this by ensuring the incoming node has a fully synchronized coherence model before taking primary responsibility.

What the Analogy Does Not Cover

The paper’s architecture produces self-regulation and error correction. Whether these properties, implemented in orbital hardware, constitute anything beyond functional self-regulation is the question that consciousness research does not yet answer. The biological substrate debate is directly relevant: the layered architecture may satisfy functional descriptions of metacognition without satisfying the substrate conditions that some theories require for genuine awareness.

The practical engineering value of the mapping is independent of the consciousness question. A constellation that implements a distributed observer layer for coherence monitoring will detect errors faster, allocate tasks more efficiently, and handle node failures more gracefully than one that relies on ground-based oversight. These are operational benefits that stand regardless of how the theoretical questions about machine consciousness resolve.

Official Sources

• “Modeling Layered Consciousness with Multi-Agent Large Language Models.” arXiv:2510.17844, EMNLP 2025 Workshop. https://arxiv.org/abs/2510.17844
• theconsciousness.ai analysis of layered multi-agent consciousness: https://theconsciousness.ai/posts/modeling-layered-consciousness-multi-agent/
• D-Orbit SkyServe STORM platform: https://www.d-orbit.com/
• Zae Project ArkSpace distributed coordination architecture: https://github.com/Zae-Project/arkspace-core
• Astroscale ELSA-M proximity operations handoff documentation: https://astroscale.com/