Engineering Specification for Artificial Consciousness¶
The Four-Model Theory states what consciousness is in architectural terms — so it doubles as a concrete specification for building one: a system that implements the four model kinds under closure, with free compute, embodiment and at least one pain-pleasure axis, which may be engineered. Building such a system is a downstream implication of the theory, not the theory itself — and at scale it remains an unrealized engineering problem.
Implementation blueprints now exist for other accounts as well: S3Q specifies a five-layer pipeline built on a world model, internal simulation and prediction–observation coherence (Grinberg et al., 2026), and CTM-AI and ReCoN-Ipsundrum are two more. S3Q derives a basic self from action–outcome linking rather than positing a separate explicit self-model, and it requires neither self-referential closure nor a Class 4 regime. Those three requirements — an explicit self-model, self-referential closure and the Class 4 regime — are what the Four-Model Theory adds. Integrated Information Theory, by contrast, defines a mathematical quantity (Phi) whose computation is intractable for any system of realistic size.
FMT is first a theory of what consciousness is, not a build manual; that it is architectural enough to read as a blueprint does not make consciousness a solved engineering task. No full-scale system has been built, and a scaffolded agent loop with external memory — however sophisticated — is architectural mimicry, not the self-referential closure the theory requires. What follows is the specification such a system would have to meet, together with the component mechanisms that can be probed in silico long before any whole is attempted.
The Specification¶
The engineering requirements are derived directly from the theory's two thresholds and from its sufficient conditions. The thresholds are each necessary and neither is sufficient; "what is sufficient is a conjunction: a system that implements the four model kinds under closure, with free compute, embodiment and at least one pain-pleasure axis, is conscious." Requirements 1–4 restate the thresholds; Requirements 5 and 6 complete the conjunction.
Requirement 1: A substrate with free compute. The computational medium must be capable of Class 4 (universal) computation and actually deploy it on ongoing self-referential modeling — free compute, not the single feedforward pass of a transformer. In the human brain the signature of this regime is near-criticality — power-law avalanches, long-range temporal correlations, a branching ratio near one — which is how the requirement is measured in brains, not a separate box to tick. Across substrates those statistics need not co-occur, so the requirement is stated as Class 4 capability rather than proximity to any single criticality measure. Candidate substrates include recurrent networks that sustain self-referential computation, neuromorphic hardware, or other architectures capable of continuous Class 4 dynamics.
Requirement 2: Four model kinds along two axes. The system must implement:
- An Implicit World Model (IWM): a substrate-level store of accumulated world knowledge, encoded in the system's connective architecture rather than in explicit representations. Learned, persistent, non-conscious.
- An Implicit Self Model (ISM): a substrate-level store of self-knowledge — the system's "body schema," operational parameters, calibration state, accumulated behavioral patterns. Also learned, persistent, non-conscious, and critically distinct from the system's explicit outputs about itself.
- An Explicit World Model (EWM): a dynamically generated, transient, virtual construction — the system's "conscious" representation of its environment. Generated from the IWM and current input. Exists only while the computation runs.
- An Explicit Self Model (ESM): a dynamically generated, transient, virtual construction — the system's ongoing self-simulation. Generated from the ISM and self-referential input. The locus of what the theory identifies as subjective experience.
Requirement 3: The real/virtual split. The implicit models (substrate-level, "real side") and explicit models (computational-level, "virtual side") must constitute genuinely distinct ontological levels. The explicit models must be generated by but not identical to the implicit models — virtual, transient, and possessing properties (including virtual qualia) that are constitutive at the computational level but incoherent at the substrate level.
Requirement 4: Self-referential closure. The ESM must model the system that is generating the ESM. This self-referential loop — the system's model includes a model of itself modeling — collapses the inside/outside distinction and is what the theory identifies as the mechanism that makes experience constitutive rather than additional. The self model must be a model of the system itself, however simple: "a realistic representation of the organism — its states, its geometry, its capabilities — in proportion to what the system has. A position estimate alone is not one." The white-box trace works through the robot vacuum that fails this test.
Requirement 5: Embodiment. The system needs a body for its self model to model, physical or a persistent artificial embodiment. "Embodiment is required for consciousness to emerge naturally, not for it to continue existing: the body is what gives the self model its content." A system with almost no body to model would have almost nothing for the self model to hold: the theory places a person locked in from birth "near the floor of minimal consciousness."
Requirement 6: At least one pain-pleasure axis. The system must have at least one valence axis along which its states are better or worse for it. The axis may be engineered: "Presetting pain as aversive and pleasure as appetitive shortcuts what selection does by eliminating the creatures that get the sign wrong, so an engineered axis is compatible with substrate independence."
What the Theory Predicts¶
If the specification is met, the theory makes a bold prediction: the resulting system would not merely imitate consciousness but would be conscious — possessing genuine phenomenal experience constituted by its virtual models. The difference between interacting with such a system and interacting with any current AI may become apparent as the interaction lengthens. Imitation is possible in principle; sustaining it without the architecture carries a cost no current system pays.
This prediction faces the other-minds problem: no behavioral test can conclusively demonstrate consciousness. The difference the theory asserts is architectural rather than impressionistic: closure, persistence and computational regime are properties measurable on the system itself, not qualities a user is asked to sense in the interaction. An artificial system's weights and running dynamics are open to inspection, so closure can be checked by the white-box trace rather than inferred from behavior.
Figure¶
graph TB
subgraph substrate["Substrate Layer (Real Side)"]
direction LR
CRIT["Free Compute<br/><i>Deployed Class-4 capability;<br/>criticality = its signature</i><br/>Recurrent, self-sustaining"]
IWM["IWM<br/><i>World knowledge in<br/>connective architecture</i>"]
ISM["ISM<br/><i>Self-knowledge in<br/>substrate parameters</i>"]
end
subgraph virtual["Virtual Layer (Phenomenal in the Closed Loop)"]
direction LR
EWM["EWM<br/><i>Generated world<br/>representation</i>"]
ESM["ESM<br/><i>Generated self-<br/>simulation</i>"]
end
CRIT -->|"sustains"| virtual
IWM -->|"generates"| EWM
ISM -->|"generates"| ESM
ESM -->|"self-referential<br/>closure"| ESM
EWM <-->|"interaction"| ESM
virtual -->|"evaluation<br/>feeds back"| substrate
style substrate fill:#264653,color:#fff,stroke:#1d3557
style virtual fill:#e76f51,color:#fff,stroke:#9b2226
style CRIT fill:#2a9d8f,color:#fff
style IWM fill:#264653,color:#fff
style ISM fill:#264653,color:#fff
style EWM fill:#e76f51,color:#fff
style ESM fill:#e76f51,color:#fffThe engineering specification in architectural form. The substrate layer must deploy free compute — of which near-criticality is the neural signature — and house two implicit models. The virtual layer — generated from, but ontologically distinct from, the substrate — hosts the explicit models where experience is constitutive. Self-referential closure (ESM modeling itself) is the mechanism that distinguishes this system from a mere simulation.
Partial Implementations¶
The specification also predicts what partial implementations would look like. Systems with some but not all components should show the component mechanisms of the parts they have and no consciousness, both thresholds being necessary:
- Four model kinds without free compute: architecture present but the self-referential computation not running — analogous to a brain under anesthesia. No consciousness.
- Free compute without the four model kinds: universal computation deployed but no self-simulation — analogous to a weather system. No consciousness.
- Two models with free compute (e.g., IWM + EWM only, no self-models): world-modeling without self-modeling. Four is the theory's minimum, so this falls short of consciousness; the prediction is the world-modeling mechanism without consciousness, detectable through behavioral signatures and computational complexity measures.
These intermediate cases offer empirically testable predictions before the full specification is achievable.
Key Takeaway¶
Because the Four-Model Theory specifies what consciousness is in architectural terms, it also yields a concrete blueprint: the four model kinds on a substrate with free compute, closed by self-referential modeling, with a body for the self model to model and at least one pain-pleasure axis, which may be engineered. Other accounts now have blueprints too; what FMT adds is an explicit self-model, self-referential closure and the Class 4 regime. Building one remains a downstream implication, not a claim that consciousness has been reduced to an engineering deliverable.