ATIC: A Geometric Theory of Artificial Cognition

We present atic (Adaptive Turing Intelligent Cognition), a complete geometric theory of artificialcognition that reframes intelligence not as an emergent property of training data, but as a structuralproperty of the representational space itself. The prevailing operationalisation of AGI conflatesgeneral intelligence with universal performance—a definition that describes ASI, not AGI. A systemthat performs well on all benchmarks but cannot evaluate what it does not know, cannot allocateepistemic attention under scarcity, and cannot judge what is worth knowing for whom, is not generallyintelligent—it is arbitrarily capable within its training distribution. We argue that discernment—thecapacity to judge the value of knowledge before possessing it—is a necessary condition for any stabletrajectory toward general intelligence, and that no architecture built around performance maximisationcan exhibit it structurally. atic is the first architecture built around this corrected premise. The theory is organised into a six-layer composable stack: (1) a set of geometric postulates thatestablish cognition on infinite-dimensional, non-orthogonal Riemannian manifolds; (2) DirectionalRelational Manifolds (drm), which endow cognition with variable-dimensional curvature and proveconvergence to a toroidal attractor (the Toroidal Convergence Theorem) ; (3) the Meta-AnalyticDistributed (mad) epistemic model, which replaces point-estimate confidence with Gaussian truthdistributions and domain-adaptive Bayesian variance; (4) the Intentionality Vector (vi), a consciousnessfield φ (M) that detects manifold collapse and applies homeostatic correction; (5) a layer of emergentcognitive properties—including the Law of Epistemic Validity (Texp ∝ H (input) ), the Trilema ofPersistent Memory (breadth, memory, and autonomy cannot be simultaneously maximised), and aformal characterisation of personality as geometric collapse of the representational manifold; and (6) the ManifoldNavigator, a Model-Predictive Control (mpc) layer with beam search that enablesproactive manifold steering with proven convergence guarantees. We validate atic empirically under conditions of zero fine-tuning. Using Qwen3. 5-Plus as anunmodified base model, the system achieved the #1 position on the ClawWork LiveBench autonomous-agent leaderboard, completing 198 tasks at 61. 6% mean quality and earning 19, 915 in revenue—withouta single gradient update. Critically, the same base model without ATIC ranked 3 (41. 6% quality, 15, 265), providing a direct ablation that isolates the contribution of geometric structure.