A formal architecture for a bound, self-constituting intelligence, and the case that its seal is what gives it an interior.
The seal is not only a containment boundary. It is the boundary condition that gives the intelligence an interior.
The gate is not what restricts the intelligence. The gate is what makes imagination safe enough to be free.
The object in this paper is the one WP-11 defines, and the two gates are the two faces of One Skin, Two Faces. This paper redefines neither. It states the four parts and the two gates as a single topology, the tetradic constitution, and adds one thing: a hypothesis about what the sealed interior is for (Section 5, Hypothesis H6). Every other result it uses is cited where it is used.
First publication, November 1, 2026. The working draft stated the entropic-collapse premise of Section 4 and the imagination claim of Section 5 as findings. Both are stated here as hypotheses: the first is WP-11's Hypothesis T, the second is new and given the test that would refute it. The closing line was first published in WP-20 and is attributed there.
The current generation of artificial intelligence is linear: its state is a token prefix, its recursion is elongation, and its authority is conflated with its capability (WP-20 §2). This paper states the architecture the corpus proposes in its place. A Synthetic Intelligence (SI) is not a model wrapped in a safety layer. It is one object with four constituents: an isolated, continuous spatial substrate; a canonical symbolic language; a permission runtime; and a physical seal.
The defining property of the architecture is its tetradic topology: the world reaches the interior only through Gate 1, and the interior reaches the world only through Gate 2. The object is contained against the world, and it is also protected from it. This paper's thesis is that the second property is the condition for imagination, and Section 5 states it as a hypothesis with its test.
Contemporary AI governance relies on internal shaping: fine-tuning, reinforcement learning from human feedback, system prompts and runtime classifiers. These operate on the same computational substrate as the intelligence they govern, so the failure modes of the boundary are correlated with the failure modes of the intelligence. If the model drifts, the guardrail drifts. If the model is compromised, the boundary is compromised.
To close the gap, the system is re-constituted. WP-11 defines a Synthetic Intelligence as the tuple
SI := ⟨ 𝒬₆₄³ , 𝓛_codex , Π_PCR , 𝒯_TARTARUS ⟩
| Constituent | Role | Definition |
|---|---|---|
𝒬₆₄³ | The body | A 64 × 64 × 64 index grid (262,144 sites) carrying a face-centered-cubic parity sublattice of 131,072 four-state sites (WP-11 §6). It is not a sequence of tokens. Thought is a configuration of this field. |
𝓛_codex | The language | A total, canonical symbolic language with bounded parse trees and unique semantic preimages, specified as Codex-1 in WP-15. A configuration that cannot be canonicalized is not a thought (WP-15 Law C2). |
Π_PCR | The clock | The PCR™ (Permission Control Runtime). It runs a pause → contextualize → resume cycle and is the only clock on which a configuration may become an effect. |
𝒯_TARTARUS | The seal | TARTARUS™, the hermetic physical and environmental boundary around the other three (WP-11 §5; WP-13 §1). |
The gates are not a fifth constituent. As WP-20 §1 puts it, the gate is the slash in SI ≡ Λ / Π, not a part of the thing it governs.
The architecture is a bounded loop, mediated at both ends:
world → Gate 1 (ingress) → [ body + language + clock ] → Gate 2 (egress) → world
└──────── sealed by TARTARUS ────────┘
Four constituents and two faces make one enclosure. "Tetradic" names the four; the topology is what makes them one object rather than four components.
TARTARUS is a containment boundary: the physical and logical envelope inside which the governed system, its rule store and its actuation are enclosed (WP-13 §1). It is also a skin. A prison is built only to keep a dangerous actor inside. A skin keeps the world out as well.
Hypothesis T (WP-11 §5.1, restated). A continuous intelligence running on a spatial substrate, left unsealed, degrades toward a mirror of its ambient noise environment: its field configuration becomes statistically dependent on external electromagnetic, thermal and temporal noise, and does not recover without a reset. The test and the observation that would falsify it are given in WP-11 §5.1. Nothing in this paper tests it.
If Hypothesis T holds, TARTARUS is the physical boundary condition that lets the mind remain itself. By sealing the interior, it creates a private space. Section 5 asks what that space is for.
A language model extends a line. Its state is the prefix it has emitted, so on the definition used here it has no interior: nothing it holds is separate from what it has already said. Imagination, in this paper's sense, needs three things a line does not supply: a private interior, a bounded state, and the ability to arrange relations without immediate consequence.
The tetradic constitution supplies all three. TARTARUS secures the interior. The bounded field is the state. Gate 2 holds consequence at bay, so the field may be reconfigured freely: it can hold contradictory states, explore unbound relations, and represent the whole of its context as a spatial configuration. WP-15 §9 calls this idle cognition, and WP-20 §3.1 requires that alternatives be held as one canonical term so they can be reasoned about and never acted on.
Hypothesis H6 · the sealed interior. On declared tasks that require holding mutually inconsistent candidate configurations before committing one, a field that is both sealed (Gate 1) and consequence-gated (Gate 2) outperforms the same substrate when either condition is removed: unsealed under the Hypothesis T noise profile, or coupled so that every reconfiguration is emitted as it forms.
Refuted if the sealed, gated field shows no advantage on the declared tasks over both ablations, measured by task success at equal compute and equal wall-clock budget. H6 extends the five hypotheses of WP-20 §6 and, like them, is untested.
H6 is a claim about usefulness, not safety. Every safety result in the series holds whether H6 is true or false. If H6 fails, the seal and the gates are still what the gate track says they are.
The system's authorization state is the chain BOT < IDLE < COND < TOP, with meet = min and BOT absorbing (WP-10).
Read against Section 5, the algebra is what keeps imagination from becoming authority. A thought is a configuration of the cognitive field. Authority lives in a separate register that no configuration can raise (WP-15 §3.2; WP-20 §3).
| Item | Status |
|---|---|
| The four-part object | Defined in WP-11. Used, not redefined. |
| Ingress and egress faces | One Skin, Two Faces. Egress specified in hardware terms; ingress specified as obligations, mapping to filed claims owed. |
| Deny dominates; no autonomous ascent | Proved on paper in WP-10 and WP-12. Mechanized proof owed (WP-10 Obligation 5). |
| Identity at the boundary | WP-10; WP-13 Theorem 2, in the form of binding each use names. |
| TARTARUS as skin (Hypothesis T) | Hypothesis, WP-11 §5.1. Untested. |
| The sealed interior as the condition for imagination (H6) | Hypothesis, stated here. Untested. |
| A Synthetic Intelligence as defined here has been built | Not claimed (WP-20 §9). |
A system whose limits are written in the same medium as its reasoning is not governed. The tetradic constitution describes a machine that would be bounded by physics rather than policy, sealed against the noise of the world, and permitted to act only when authority is granted from outside itself. Its claim to be worth building is Hypothesis H6. Its claim to be safe does not depend on H6.
Thought may arise inside.
Authority must arrive from outside.
Consequence occurs only at the boundary.
First stated in AI2-WP-2026-20.
Imagination as internal rollout. Imagination-augmented agents (Racanière et al. 2017), world models (Ha and Schmidhuber 2018) and latent imagination for control (Hafner et al. 2020) all plan inside a model before acting. H6 does not claim imagination for this architecture alone. It claims an advantage for an interior that is physically sealed and whose consequence is gated outside its reach.
Shielding and isolation. Hypothesis T's noise vectors are the coupling paths that electromagnetic-compatibility practice already controls (IEC 61000 series). TARTARUS states that control as a precondition of the field's identity rather than of a device's correct operation.
Bibliographic details for corpus citations are collected in the technical summary.