Anisotropic Compute as a First-Class Architectural Principle: Integrating Directional Asymmetry into the Janus-Class AI Processor

Anisotropic Compute as a First-Class Architectural Principle: Integrating Directional Asymmetry into the Janus-Class AI Processor Abstract Contemporary artificial intelligence processors are predominantly designed under the assumption of isotropy: the premise that computational rules, data movement costs, and execution semantics should remain uniform across the spatial and temporal dimensions of the die. While this symmetric approach facilitated the rapid scaling of the von Neumann and Harvard architectures during the Moore's Law era, it increasingly conflicts with the physical, energetic, and stability constraints encountered at the exascale. The current trajectory of "brute-force" scaling—characterized by ever-wider buses, higher clock frequencies, and massive homogeneous parallelism—has yielded systems that are thermodynamically fragile and prone to runaway feedback loops. As AI workloads transition from static inference to autonomous, long-horizon agentic behaviors, the requirement for architectural stability becomes paramount, necessitating a fundamental departure from isotropic design philosophies. This white paper formally introduces Anisotropic Compute as a first-class architectural principle within the Janus-Class AI Processor. The principle mandates the intentional, structural separation of high-velocity information transport from high-inertia state evolution, embedding constraint-aware asymmetry directly into the computational substrate. Rather than relying on exotic materials or speculative future physics for implementation, this work establishes anisotropy as a design law applicable across logical, spatial, and physical layers of standard CMOS and chiplet-based systems. By orthogonalizing the axes of growth and control, the architecture allows for the coexistence of extremely low-latency execution paths with extremely high-assurance governance mechanisms—two requirements that are mutually exclusive in isotropic topologies. Significantly, this architectural thesis draws independent validation from recent experimental observations in condensed-matter physics, specifically the discovery of direction-dependent quasiparticle behavior (semi-Dirac fermions) in zirconium silicon sulfide (ZrSiS). These physical systems demonstrate that nature itself resolves the trade-off between propagation speed and structural stability through geometric asymmetry—effectively behaving as massless along one axis and massive along the perpendicular. The Janus-Class architecture adopts this same constraint geometry to orthogonalize the functions of growth (Adaptive Mode) and control (Reflective Mode), ensuring that AI systems can scale capability without sacrificing the mathematical guarantees of stability required for autonomous operation. 1. Motivation: The Limits of Isotropic Compute 1.1 The Scaling Failure Mode: Entropy and Data Movement The foundational crisis of modern high-performance computing (HPC) and AI hardware is not a lack of arithmetic logic units (ALUs), but a fundamental failure of transport thermodynamics. Modern AI workloads are no longer compute-bound; they are strictly data-movement bound. The energy cost of moving a datum across a 7nm chip is orders of magnitude higher than the cost of performing a floating-point operation on that datum.1 As process nodes shrink, the relative cost of wire delay and interconnect power dissipation has skyrocketed, creating a regime where the computational core is effectively held hostage by its own communication infrastructure. In the prevailing "Isotropic" design paradigm—exemplified by standard Network-on-Chip (NoC) mesh topologies and symmetric multicore processors—the architecture assumes that data should be able to move in any direction with equal facility. This assumption, while simplifying the logical design of the scheduler, results in a catastrophic mismatch with physical reality at scale. Isotropic architectures attempt to solve latency and bandwidth bottlenecks through brute-force mechanisms: wider buses, deeper buffers, complex coherency protocols, and massive energy dissipation in the interconnects.2 This approach yields several critical failure modes: Uniform Latency Expectations: The system attempts to optimize every path for speed. This leads to excessive buffering and synchronization overhead as the system tries to maintain a fiction of uniform access times across a physically distributed substrate. Thermal Homogeneity: Because activity is allowed to propagate uniformly, "hot spots" migrate unpredictably across the die. This stochastic thermal behavior makes targeted cooling impossible, forcing the entire package to be throttled by the worst-case localized thermal event. Runaway Feedback Loops: In deep learning training and, more critically, in agentic workflows, the lack of structural "damping" allows errors or hallucinations to propagate as rapidly as valid signals. An isotropic substrate offers no physical resistance to logical instability; a wrong signal travels just as fast as a right one.3 As described in the analysis of the CollectiveOS ecosystem, the industrial model attempts to mitigate these physical limits through the "Brute Force" application of energy and capital—essentially fighting the laws of thermodynamics. This manifests in the "Trillionaire Trajectory" of building ever-larger data centers with liquid cooling rivers, a solution that scales asymptotically toward failure. Just as the Tsiolkovsky Rocket Equation limits the viability of massive space-based solutions by imposing exponential fuel costs on mass 3, Landauer’s Principle and interconnect latency limits the viability of massive isotropic silicon. The energy required to erase information (resetting the state) and the energy required to transport it eventually exceeds the value of the computation itself.4 1.2 Stability as an Architectural Requirement The transition of artificial intelligence from passive "chatbots" to active "agents" marks a critical inflection point in architectural requirements. When an AI system acts as a passive oracle (e.g., GPT-4 responding to a prompt), instability is a nuisance—a hallucination or a timeout. However, when an AI system acts as an autonomous agent (e.g., the Janus/CollectiveOS "Rabbit" agent executing scripts, managing resources, or controlling robotics), stability becomes a safety-critical requirement.3 Current architectures rely almost exclusively on software-layer "guardrails" to prevent erratic behavior. These include RLHF (Reinforcement Learning from Human Feedback), constitution files, and prompt engineering. However, if the underlying hardware substrate encourages uncontrolled signal propagation, software constraints are fighting an uphill battle against the physics of the chip. A system designed to maximize throughput at all costs will naturally resist constraints that impose latency or checks. Stability must therefore be enforced at the architectural level. This requires a shift from Isotropic designs, where all directions are equal, to Anisotropic designs, where the hardware itself imposes a "grain" or directionality that favors stability over speed in specific, critical dimensions. This aligns with the Universal Intent Layer (UIL) framework, which posits that stable systems are not random accidents but the result of adhering to deep informational constraints.6 Just as biological systems use structural anisotropy—such as the orientation of myofibrils in muscle tissue to direct force, or the polarity of axons to direct information—a resilient AI processor must utilize structural anisotropy to direct information productively. The "Universal Intent Layer" (UIL) suggests that the universe organizes itself according to deep informational constraints, formalized by the inequality $P(X | UIL) \gg P(X | \text{random})$. This implies that stable, ordered states are statistically probable only when constrained by lawful attractors. In the context of chip design, this means that "random" access (isotropy) is inherently less stable than "constrained" access (anisotropy).3 2. The Anisotropic Compute Principle 2.1 Definition and Core Axiom Anisotropic Compute is defined as the intentional design of computational systems in which different operational directions—logical, spatial, or temporal—obey fundamentally different physical and informational constraints. It is the rejection of the "Sphere of Uniformity" in favor of the "Cylinder of Flow." In an isotropic system, a packet moving from Core A to Core B incurs roughly the same "physics" (latency, energy, verification cost) regardless of the content of that packet. In an anisotropic system: Transport $\neq$ Retention: The axis optimized for the movement of information is structurally distinct from the axis optimized for the retention and verification of state. Growth $\neq$ Control: The pathways allowing for system expansion (throughput, exploration) are orthogonal to the pathways enforcing system boundaries (governance, constraints). The Core Axiom of Anisotropic Compute: High-velocity information transport and high-inertia state evolution must never share the same dominant axis. This axiom is the hardware realization of the "Separation of Mechanism and Policy" principle, a foundational concept in secure operating systems.7 By physically orthogonalizing these functions, the architecture prevents high-speed "hot" data from corrupting the "cold" stable state, and prevents heavy governance logic from choking the high-speed data path. 2.2 The Physics of Asymmetry The necessity of this principle is grounded in the physics of inertia. In mechanical systems, high inertia is required to maintain stability at high velocities, yet high inertia makes acceleration difficult.8 An object with high mass resists perturbation but requires enormous energy to move. An object with low mass moves easily but is easily knocked off course. An isotropic system tries to be both "light" (for speed) and "heavy" (for stability) everywhere at once. It attempts to make memory access fast (low inertia) while making memory retention stable (high inertia). This leads to the compromise of complex cache hierarchies that are neither perfectly fast nor perfectly stable, consuming vast amounts of power to maintain coherency. Anisotropic design resolves this by spatially separating the "mass." One direction is "massless" (low inertia, high velocity) to facilitate rapid signal propagation and execution. The perpendicular direction is "massive" (high inertia, low velocity) to facilitate robust state retention, error correction, and governance. This allows the system to possess "Variable Inertia" depending on the vector of operation.8 3. Janus-Class Context 3.1 The Janus Dual-Mode Logic: Temporal Anisotropy The Janus-Class processor, as detailed in the CollectiveOS specifications, is already logically anisotropic in the temporal domain. It operates under two distinct regimes governed by the Living Fibonacci Engine (LFE).3 The LFE is a biomimetic adaptive controller that modulates the system's operational phase based on a stability metric derived from the Golden Ratio ($\Phi \approx 1.618$). The system operates in two modes: Adaptive Mode ($c=+1$): The "Forward Face." This mode prioritizes growth, exploration, and throughput. It corresponds to the "Massless" axis—rapid expansion of the operational context, high-frequency token generation, and aggressive resource consumption. In this mode, the system behaves like a standard Fibonacci growth sequence, maximizing $R_n$ (growth ratio). Reflective Mode ($c=-1$): The "Backward Face." This mode prioritizes consolidation, repair, and constraint enforcement. It corresponds to the "Massive" axis—slowing down to validate data against the Proof Vault, optimizing memory structures, and dissipating accumulated entropy (heat/error). This mode is triggered when the "Golden Error" ($\epsilon_n$) exceeds a safety threshold. The mathematical operation of the LFE ensures that the system does not spiral into chaotic oscillation. The phase space trajectory of the engine reveals a "breathing" cycle: the system expands in Adaptive Mode following the Golden Ratio growth vector, spiraling outward. However, upon exceeding the Golden Error threshold, the controller flips the parameter $c$ to -1. This induces a damping orbit—a "snap back" to stability—that returns the state to the stable manifold. This cyclical interplay prevents the integral windup common in standard PID controllers, effectively "metabolizing" instability rather than fighting it.3 The Anisotropic Compute principle provides the spatial counterpart to this temporal duality. It ensures that when the LFE switches modes, the hardware naturally supports the required physics. Adaptive processing flows along the low-resistance transport planes, while Reflective processing engages the high-resistance governance planes. 3.2 Orthogonalization of Function within CollectiveOS The CollectiveOS architecture inherently demands this separation. The system distinguishes between Compute (stateless agents like "Rabbit") and State (the text brain/God File).3 The "Rabbit" Axis (Execution): Rabbit requires high-velocity, low-latency access to temporary registers and inference engines to execute scripts. This is the Data Plane. Rabbit operates in the "Now," processing streams of tokens and sensor data. The "Syn" Axis (Memory/Truth): Syn requires high-integrity, high-inertia access to the immutable logs of the Proof Vault. Latency is secondary to immutability and verification. This is the Control Plane. Syn operates in "Eternity," managing the permanent record of the system. In a Janus-Class processor, these are not just different software threads sharing the same L3 cache; they are physically different "lanes" on the chip, optimized for their specific thermodynamic requirements. The Data Plane might use low-voltage, high-speed signaling with minimal error correction for transient data, while the Control Plane uses higher-voltage, robust signaling with heavy ECC and cryptographic verification for persistent state. 4. Independent Physical Validation: The Semi-Dirac Bridge The assertion that speed and stability can be reconciled through directional asymmetry is not merely a computational convenience; it is a physical reality recently observed in quantum materials. The behavior of semi-Dirac fermions in zirconium silicon sulfide (ZrSiS) provides a striking "existence proof" for the Janus-Class architectural philosophy.10 4.1 Direction-Dependent Dispersion in Condensed Matter In 2024, researchers confirmed the existence of quasiparticles in ZrSiS that exhibit a hybrid dispersion relation. These particles behave as massless Dirac fermions along one crystallographic direction and as massive Schrödinger fermions along the perpendicular direction.12 Direction A (The Massless Axis): Along this axis, the energy-momentum relationship is linear ($E \propto k$). The effective mass is zero ($m^* \approx 0$). Charge carriers move with extreme velocity, similar to photons. This allows for highly efficient, dissipation-less transport of information. This is the physical analogue of the High-Velocity Transport Plane in the Janus processor. Direction B (The Massive Axis): Along the orthogonal axis, the dispersion is quadratic ($E \propto k^2$). The particles exhibit significant effective mass ($m^* > 0$). They possess inertia and are less susceptible to perturbations, behaving like conventional electrons. This provides the "weight" necessary for localization and interaction. This is the physical analogue of the High-Inertia State/Governance Plane in the Janus processor. The scaling of Landau levels in these materials follows a unique $B^{2/3}$ power law.11 This is distinct from standard massive particles (which scale linearly with magnetic field $B$) and massless particles (which scale with $\sqrt{B}$). This unique $B^{2/3}$ scaling indicates a new regime of physics that is "critical" in one dimension and "stable" in the other—exactly the dual property required for a self-stabilizing AI architecture. 4.2 Relevance to Architecture (Conceptual Only) It is critical to clarify that the Janus-Class architecture does not propose using ZrSiS or exotic topological semi-metals as the fabrication material for the processor in the near term. Rather, the discovery of semi-Dirac fermions serves as independent validation of the constraint geometry. Nature has demonstrated that "mass" (inertia/resistance to change) need not be a scalar property of an object; it can be a tensor property of the space the object moves through.16 A particle can be "light" when it needs to move and "heavy" when it needs to stay put, simply by changing direction. The Janus-Class processor implements this logic in silicon: Logical Anisotropy: Packets moving "North-South" (Data) encounter minimal switching overhead, no deep inspection, and high bandwidth (Massless). Logical Inertia: Packets moving "East-West" (Control) encounter deep inspection, logging, cryptographic signing, and validation overhead (Massive). This alignment with physical law suggests that Anisotropic Compute is not an arbitrary engineering choice, but a convergence toward a fundamental principle of efficient systems. Systems that survive in nature are those that can move fast to acquire resources but stay stable to retain them; the semi-Dirac fermion is the quantum mechanical embodiment of this survival strategy. 5. Architectural Implications The implementation of Anisotropic Compute in the Janus-Class processor manifests as a rigorous separation of the Data Transport Plane and the State & Governance Plane. This is not merely a logical separation (as in SDN control planes 17) but a physical one, influencing the interconnect topology, memory hierarchy, and power distribution. 5.1 Data Transport Planes (The Massless Axis) The Data Transport Planes are designed for Adaptive Mode operations. They are optimized for the high-velocity propagation of transient signals—inference tokens, immediate robotic actuator commands, and "Rabbit" agent scratchpad operations. Topology: These planes utilize low-diameter, high-radix topologies (e.g., flattened butterfly or dragonfly networks) that minimize hop counts.18 The goal is to make the network appear as "flat" as possible to the executing agent. Protocol: Transport is handled via circuit-switched or wormhole routing mechanisms that establish a pre-reserved path, effectively removing the "mass" of routing decisions from the packet's journey.19 Once a path is established, data flows with minimal buffering. State: These planes are explicitly stateless. They do not contain caches or buffers for long-term retention. Data is "in flight" or it is discarded. This prevents the accumulation of "technical debt" or entropy within the fast path. Privilege: This plane has Execution Privilege but no Write Privilege to the core state. It can calculate, but it cannot permanently remember. Analogy: This corresponds to the linear, massless direction of the semi-Dirac fermion. Resistance is minimized to near-zero to allow speed-of-light (on chip) propagation. 5.2 State & Governance Planes (The Massive Axis) The State & Governance Planes are designed for Reflective Mode operations. They are optimized for the high-inertia retention of truth—the "Syn" memory, the Proof Vault logs, and the GATA PRIME policy enforcement engines. Topology: These planes utilize hierarchical, tree-based topologies that enforce a strict chain of command. Every write operation must traverse up the tree to a validation node before being committed to the immutable ledger. Protocol: Communication here is packet-switched with deep inspection. Every packet is subject to policy checks (Policy-as-Code), cryptographic hashing, and causal validation.3 This introduces latency—"mass"—which is a feature, not a bug. It dampens high-frequency oscillations and prevents "flash crashes" or hallucination cascades from becoming permanent state. State: This is the domain of Non-Volatile Memory (NVM) and WORM (Write-Once-Read-Many) storage. It is the "heavy" anchor of the system. Privilege: This plane has Write Privilege but operates at a fraction of the clock speed of the Data Plane. It acts as the "Super-Ego" to the Data Plane's "Id." Analogy: This corresponds to the quadratic, massive direction of the semi-Dirac fermion. The effective mass is high, requiring significant "force" (computational work/proof) to change the state. 5.3 Failure Containment and Systemic Resilience By constructing the architecture in this orthogonal manner, failure containment becomes an emergent property rather than an active intervention. In an Isotropic system, a runaway thread (e.g., a hallucinating agent) consumes bandwidth and memory indiscriminately, potentially crashing the entire node by flooding the coherence bus. In the Janus Anisotropic system, the failure mode is physically constrained: The runaway thread accelerates along the Massless Axis (Transport), generating a storm of tokens. However, to affect the world or the system state, these requests must cross into the Massive Axis (Governance) to be logged or actuated. The Massive Axis has high inertia. It imposes a "speed limit" defined by the cryptographic validation rate. The high-velocity stream "impacts" this heavy boundary and is dissipated. The Governance plane simply rejects the excess transactions that do not have valid cryptographic signatures or that violate LTL (Linear Temporal Logic) safety constraints.3 The Living Fibonacci Engine detects this "Golden Error" (high deviation in resource consumption vs. valid state changes) and triggers a Reflective Mode switch. This switch clamps down on the Transport plane's resources, effectively "freezing" the massless axis until stability is restored. Stability does not rely on a global halt or a complex software watchdog; it relies on the physical inability of "light" errors to move "heavy" truth. 6. Relationship to the Living Fibonacci Engine (LFE) The Living Fibonacci Engine and Anisotropic Compute are the temporal and spatial expressions of the same control logic. They form a unified homeostatic loop. The LFE provides the Temporal Constraint: It answers when the system should expand or consolidate. It monitors the $\epsilon_n$ metric and flips the $c$ parameter. Anisotropic Compute provides the Spatial Constraint: It answers where expansion and consolidation occur. When the LFE is in Adaptive Mode ($c=+1$), the processor actively power-gates the non-essential parts of the Governance Plane (reducing its drag) and overclocks the Transport Plane. The system becomes "lighter" and more agile, capable of rapid exploration. When the LFE shifts to Reflective Mode ($c=-1$), the processor throttles the Transport Plane (increasing resistance) and fully powers the Governance/Validation Plane. The system becomes "heavier" and more rigid, prioritizing safety and repair. This dynamic modulation of the system's "effective mass" allows the Janus processor to navigate chaotic environments with the agility of a biological organism, rather than the rigidity of a machine. It is a "breathing" architecture, expanding and contracting in response to stress, guided by the mathematics of the Golden Ratio. 7. Implementation Neutrality A critical requirement of the Janus-Class architecture is that it must be buildable now, using existing semiconductor supply chains. Anisotropic Compute is a design principle, not a materials science dependency. It does not require the invention of new transistors, only the reorganization of how they are connected. 7.1 CMOS-Compatible Realization The principle applies directly to standard CMOS (Complementary Metal-Oxide-Semiconductor) processes: Asymmetric Metal Layers: Modern chips use different metal layers for interconnects (M1-M12+). Upper layers are thick, sparse, and fast (low resistance); lower layers are thin, dense, and slow. Anisotropy can be implemented by dedicating upper layers to the Transport Plane (long-distance, fast, massless) and lower layers to the Governance Plane (local, dense, massive). Heterogeneous Cores: The "big.LITTLE" architecture in mobile processors is a crude form of anisotropy. Janus refines this by specializing cores not just for power/performance, but for Verification vs. Execution. "Rabbit" cores are simple, fast RISC-V pipelines optimized for throughput. "Syn/GATA" cores are secure, out-of-order, crypto-accelerated units optimized for logic and verification. 7.2 Chiplet and 2.5D/3D Integration The rise of chiplet architectures 20 offers the perfect substrate for anisotropy. The physical separation of dies allows for distinct manufacturing processes for distinct planes. Base Die (Active Interposer): This can serve as the Massive Axis. It is built on a mature, stable node (e.g., 28nm or 14nm). It contains the slow, stable, high-capacity interconnection network, the Proof Vault storage controllers, and the LFE logic. It is the "heavy" foundation. Compute Chiplets: These are placed on top and serve as the Massless Axis. They are built on cutting-edge nodes (e.g., 3nm). They contain high-speed specialized compute (NPUs/TPUs) that communicate rapidly with each other. However, to commit state or change policy, they must communicate "down" to the base die, traversing the "gravity well" of the vertical interconnects (TSVs). This physical separation (Vertical vs. Horizontal) perfectly maps to the semi-Dirac dispersion geometry: Massless in the plane of the chiplet, Massive in the vertical connection to the base. 8. Why This Matters for AI Specifically Artificial Intelligence workloads have a unique profile that distinguishes them from traditional scientific computing or database management. This profile makes them uniquely suited for—and in need of—Anisotropic Compute. Tolerance for Approximation: AI inference is probabilistic. It tolerates "lossy" transport and approximate math (e.g., FP8 or INT4 precision). This allows the Transport Plane to be aggressively optimized for speed, shedding the weight of error correction codes (ECC) or strict coherency for transient signals. We can allow the "Rabbit" to run fast and loose because its outputs are not yet truth. Intolerance for Instability: While individual inferences can be approximate, the behavioral trajectory of an agent must be strictly bounded. A robot cannot "approximately" stop at a cliff edge; a financial agent cannot "approximately" verify a transaction. This requires the Governance Plane to be absolute, rigid, and "heavy." Isotropic architectures force a compromise: they either make the inference path too slow (by imposing safety checks everywhere) or the safety path too weak (by optimizing for speed everywhere). Anisotropic Compute aligns the hardware with the reality of AI: Loose execution, tight governance. It allows the AI to "dream" at the speed of light along one axis, but requires it to "act" with the weight of law along the other. 9. What This Architecture Is Not To maintain credibility and avoid the "hype cycle" failure mode often associated with novel AI hardware proposals, we must clarify the boundaries of this proposal: No New Physics: We are not claiming to manufacture chips out of ZrSiS or create actual semi-Dirac fermions in the processor. We are using the physics of these particles as a valid, experimentally verified model for architectural constraint. The implementation relies on standard electrons in standard copper wires. No Universal Acceleration: This architecture does not make all computation faster. In fact, it intentionally makes "state changes" slower (higher inertia) to ensure stability. It trades raw IOPS (Input/Output Operations Per Second) for Verified IOPS. It sacrifices the illusion of speed for the reality of stability. No "Magic" Logic: The LFE and Anisotropy do not solve the "Halting Problem" or create AGI. They simply provide a stable, bounded substrate for complex software agents to operate without self-destructing. They prevent the "Grey Goo" scenario of runaway digital consumption. 10. Conclusion The addition of Anisotropic Compute as a first-class principle strengthens the Janus-Class AI Processor by embedding stability directly into its architecture. Rather than fighting physical constraints through scale—the "Industrial" model of fighting gravity with rocket fuel—the Janus system aligns with constraints through structure. It recognizes that in a finite universe, infinite isotropic growth is impossible, but directed anisotropic flow is sustainable. Recent independent discoveries in condensed-matter physics, specifically the observation of semi-Dirac fermions in ZrSiS, demonstrate that nature itself resolves the speed-versus-control trade-off through directional asymmetry. The universe allows particles to be massless (fast) only if they are constrained to be massive (stable) in a perpendicular direction. The Janus-Class architecture arrives at this same solution from the computational side: splitting the universe of data into a "Massless" axis of transport and a "Massive" axis of truth. This convergence is not coincidental. It is the signature of a stable design. By orthogonalizing growth and control, the Janus-Class processor provides the first hardware substrate capable of supporting the "Sovereign Engineering" requirements of the CollectiveOS: systems that are powerful enough to build worlds, yet stable enough to inhabit them. Appendix A — Design Summary Growth $\neq$ Control: Separation of concerns is physical, not just logical. Speed $\neq$ Stability: High velocity requires high inertia (mass) to control; they must be on orthogonal axes. Constraints belong in structure: Software patches cannot fix hardware that encourages entropy. Appendix B — Safety & Scope Declaration No new physics claimed: Architectural biomimicry only. No materials specified: Compatible with standard silicon/chiplets. Design principle only: Implementation is vendor-neutral. Works cited RRAM Crossbar-Based In-Memory Computation of Anisotropic Filters for Image Preprocessing - AIR Unimi, accessed December 16, 2025, https://air.unimi.it/bitstream/2434/753885/2/09122505%20%286%29.pdf Networks-on-Chip in Emerging Interconnect Paradigms: Advantages and Challenges - Columbia CS, accessed December 16, 2025, http://www.cs.columbia.edu/~luca/research/eip_NOCS09.pdf Janus AI Processor Analysis.pdf Infodynamics, a Review - Article (Preprint v4) by Klaus Jaffe | Qeios, accessed December 16, 2025, https://www.qeios.com/read/2RBRWN.4 Shortcuts to Thermodynamic Computing: The Cost of Fast and Faithful Information Processing - PMC - NIH, accessed December 16, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC8960662/ A Synthetic Human + AI Architecture Aligned to the Same Constraints That Structure the Universe Itself - Zenodo, accessed December 16, 2025, https://zenodo.org/records/17682670 Separation of mechanism and policy - Wikipedia, accessed December 16, 2025, https://en.wikipedia.org/wiki/Separation_of_mechanism_and_policy Rotational system with variable inertia - DiVA portal, accessed December 16, 2025, https://www.diva-portal.org/smash/get/diva2:1671606/FULLTEXT01.pdf THE END OF THE GPU DATA CENTER A Metabolic, Mathematical, and Constraint-Governed Architecture for Sustainable AI Infrastructure - Zenodo, accessed December 16, 2025, https://zenodo.org/records/17850341 Semi-Dirac fermion - Wikipedia, accessed December 16, 2025, https://en.wikipedia.org/wiki/Semi-Dirac_fermion Quasiparticles become massless – but only when they're moving in the right direction, accessed December 16, 2025, https://physicsworld.com/a/quasiparticles-become-massless-but-only-when-theyre-moving-in-the-right-direction/ Particle that only has mass when moving in one direction observed for first time, accessed December 16, 2025, https://science.psu.edu/news/particle-only-has-mass-when-moving-in-one-direction-observed-first-time Semi-Dirac Fermions in a Topological Metal - MagLab, accessed December 16, 2025, https://nationalmaglab.org/user-facilities/dc-field/research/science-highlights/fermions-in-zirconium-silicon-sulfide/ Semi-Dirac electronic states in anisotropic stacked graphene structure - ResearchGate, accessed December 16, 2025, https://www.researchgate.net/publication/395835939_Semi-Dirac_electronic_states_in_anisotropic_stacked_graphene_structure Semi-Dirac Fermions in a Topological Metal (Journal Article) | OSTI.GOV, accessed December 16, 2025, https://www.osti.gov/pages/biblio/2479931 Bridging Geometry and Physics: Demystifying Semi-Dirac Fermions with Fundamental Density Theory | by Manuel Alfaro | Medium, accessed December 16, 2025, https://medium.com/@m.alfaro.007/bridging-geometry-and-physics-demystifying-semi-dirac-fermions-with-fundamental-density-theory-7d1904425ceb A Software-Defined Architecture for Integrating Heterogeneous Space and Ground Networks, accessed December 16, 2025, https://www.frontiersin.org/journals/communications-and-networks/articles/10.3389/frcmn.2021.717476/full One Protocol to Rule Them All: Wireless Network- on-Chip using Deep Reinforcement Learning - USENIX, accessed December 16, 2025, https://www.usenix.org/system/files/nsdi21-jog.pdf Developing Deadlock-Free Routing Algorithms in Torus NoC: A Formal Approach, accessed December 16, 2025, https://www.researchgate.net/publication/394954544_Developing_Deadlock-Free_Routing_Algorithms_in_Torus_NoC_A_Formal_Approach (PDF) ACRIN: Asymmetric Compression‐Recovery Interconnection Network for Inter‐Chip Sparsity Architecture - ResearchGate, accessed December 16, 2025, https://www.researchgate.net/publication/392814342_ACRIN_Asymmetric_Compression-Recovery_Interconnection_Network_for_Inter-Chip_Sparsity_Architecture