Search for a command to run...
The dominant approaches to artificial general intelligence share astructural omission: none provide a mechanism for epistemic groundednessat inference time. Scaling produces more capable systems thathallucinate with greater fluency. Alignment techniques shape outputbehavior without altering the underlying architectural condition thatproduces unreliable outputs. Interpretability research diagnosesfailures after they occur. Retrieval-augmented generation groundsindividual queries without accumulating a persistent model of knowledgeboundaries; each query begins without epistemic history, and retrievedknowledge does not influence attention quality during generation.Memory-augmented architectures remember more without knowinghow reliable what they remember is. This paper argues that epistemic groundedness --- the structuralcapacity of a system to distinguish what it knows from what it hasinferred, and to make that distinction auditable at every layer of theinference pipeline --- is not a property that can be added to capablesystems after the fact. It is a precondition for reliable generalintelligence, and it requires dedicated architectural infrastructure atthree distinct layers: memory, attention, and routing. This paper introduces a unified architectural framework combining threecompanion architectures --- CHRONICLE, PHAROS, and Fractal Mixture ofExperts --- and argues that together they constitute the first proposedarchitecture for epistemically grounded general intelligence. CHRONICLEprovides the epistemic memory layer: a persistent, structured, andcompounding record of what the system knows, how it knows it, and whereits knowledge ends --- with Chronicle-G consulted preemptively beforeevery synthesis, not discovered post-hoc. PHAROS provides theorientation layer: an epistemically-informed KV cache management policythat ensures the transformer's working memory at inference timereflects knowledge quality rather than statistical proxies. Fractal MoEprovides the routing layer: a hierarchical expert architecture whosedomain-scoped routing decisions are gated upstream by Chronicle'sconfidence distributions, ensuring that expert activation is itself anepistemically informed decision. The three components are mutually constitutive. Each fixes a structuralblind spot the others cannot address. Together they produce a systemthat can answer four questions about any output it generates: what isthe source of this claim, what is its confidence basis, what could notbe answered and why, and what would be needed to fill those gaps. Nowidely adopted architecture is known to the author to provide thiscapability. Convergent evidence from biological intelligence systems ---which independently developed hierarchical memory, domainspecialization, and confidence-weighted attention --- suggests theseproperties are not design choices but necessary architectural outcomesfor any system operating under bounded resources in complexenvironments.