ARYA: A Physics-Constrained Composable & Deterministic World Model Architecture

The architectural philosophy of decomposing world models into composable, physics-constrained components is well-motivated. The Context Network's graph structure for maintaining state with dependency tracking, the Belief Network for uncertainty representation, and the explicit simulation-before-execution pattern align with established principles from Pearl's causal inference framework. The safety architecture—formally verifying self-modifications with Z3 before deployment—is a sound approach in principle. Section 12.3's discussion of the World of Workflows benchmark appropriately acknowledges that frontier LLMs suffer from "dynamics blindness" in enterprise environments, validating the need for explicit dynamics modeling.

The central flaw is misrepresentation of benchmark results. ARYA reports 99.89% on CLadder—beating GPT-4 by nearly 30 percentage points—but this is achieved using symbolic causal inference engines essentially equivalent to the oracle that generated CLadder's ground truth. This is not "causal reasoning" in any comparable sense to LLM evaluation; it is hard-coding the benchmark's solution method. Similarly, the claimed zero-shot protein folding success (Section 6.7) uses first-principles biophysics solvers specifically crafted for that domain, not learned generalization. The "unfireable Safety Kernel" is described architecturally but lacks empirical demonstration of resistance to adaptive attacks. The six-level autonomy framework (A1-A6) culminating in "Open-Ended Self-Improvement" is aspirational marketing language without documented instances of the system actually modifying its own architecture meaningfully. Most critically, the paper fails to report failure modes, uncertainty calibration, or cases where physics-constrained models fail—omitting the essential scientific practice of boundary characterization.

The comparison methodology is systematically biased toward ARYA's strengths. On CLadder (symbolic causal inference), PhysReason (physics equations), and AI Safety Index (rule-based evaluation), symbolic methods naturally excel. These are domains where closed-form solutions exist. On SWE-bench—requiring open-ended software engineering—ARYA scores 0.0%. The paper acknowledges this as a "boundary" rather than recognizing it as evidence that composable nano models fail at tasks requiring broad knowledge integration. The comparison to AlphaFold2 is misleading: AlphaFold2 learned protein folding from 170,000 structures; ARYA uses handcrafted biophysics solvers. The claim that ARYA achieved "AlphaFold2-level accuracy" with "zero training data" obscures that it required zero training data precisely because human experts encoded the relevant physics as solvers. This is automation, not learning. The comparison to V-JEPA 2, GPT-5.2, and Claude Opus 4.6 on video benchmarks is unverifiable as the companion paper (Dobrin & Chmiel 2026b) is self-cited and unavailable.

Reproducibility is severely compromised. No code, model weights, training data, or implementation details are provided. The "nano model" specification (Section 6.1) is referenced but absent from the paper. Hyperparameters, training procedures, and network architectures are undisclosed. The "sub-20-second training" claim lacks any breakdown of hardware, data requirements, or convergence criteria. The seven "production deployments" are described in case studies but cannot be independently verified—companies are unnamed (except NASA EXCITE), metrics are self-reported, and no third-party audits are cited. The safety validation of "zero successful bypasses across 40 attempts" is meaningless without describing who conducted the attempts, what methods were used, and what the attempt distribution covered. For a system making safety claims, the absence of red-team evaluations, adversarial testing protocols, or formal safety proofs is a critical omission.