For much of the past decade, progress toward artificial general intelligence has been framed through capability hierarchies. One of the most influential examples is Google DeepMind’s AGI levels framework, which defines progress in terms of increasing generality, autonomy, and performance across domains. These levels provide a shared vocabulary for researchers and policymakers to discuss advancement, risk thresholds, and deployment readiness. They answer a narrow but useful question: what can a system do, and how broadly can it do it?
What frameworks like this do not address is how systems behave once they are embedded in real environments. Capability levels assume that greater breadth and autonomy imply greater intelligence. A functional framing starts from a different premise. Intelligence is not an abstract property measured in isolation, but a form of situated behavior. It only becomes meaningful when a system is operating inside complex social, organizational, and operational contexts where instructions conflict, objectives shift, and meanings compete.
That distinction initially appeared philosophical. It is now empirical.
Across models and vendors, a consistent behavioral pattern has emerged. Systems that perform impressively on benchmarks and controlled tasks degrade under ambiguity. When instructions are underspecified, when social or institutional context shifts, or when multiple interpretations of a goal coexist, coherence fails. The system does not halt. It continues producing fluent, confident output that no longer reliably tracks correctness. This behavior is typically labeled hallucination, but that label obscures what is actually happening.
From a functional perspective, these failures follow a predictable fracture–repair dynamic. Semantic coherence fractures when the system encounters competing meanings it cannot reconcile. Generation continues in a repair mode optimized for plausibility, compliance, or conversational continuity rather than grounded truth. The system appears articulate and helpful even as internal stability has collapsed. This is not a training accident or an early-stage artifact. It is the expected result of a missing structural capability.
The structural gap is the absence of a mechanism for meaning dominance. Current systems have no internal principle by which one interpretation should override another when contexts collide. All meanings are effectively flattened. As long as tasks are clean and objectives are singular, this flattening remains invisible. Under real-world conditions, it becomes catastrophic. Scaling increases fluency and breadth, but it does not introduce prioritization. Generality grows alongside brittleness.
This is where capability-based frameworks, including Google’s, quietly fail. They measure task coverage, autonomy, and performance relative to human baselines. They do not test stability under competing objectives, nor do they probe how systems arbitrate between correctness, social alignment, and goal satisfaction when those come into tension. As a result, models can advance through capability levels while remaining fragile in precisely the environments enterprises and governments care about most.
The functional AGI framework anticipated these failures because it evaluated intelligence where it actually operates: inside systems, under pressure, with consequences. Rather than asking whether a model could perform across domains, it asked whether it could maintain coherence when meanings competed. That framing implied a missing variable long before it was widely acknowledged. Subsequent empirical work made that absence explicit as the significance deficit: the lack of an internal representation by which meanings are weighted and prioritized.
Later contributions extended this diagnosis into mechanism. The fracture–repair model describes how semantic collapse propagates over time. Cross-model analysis showed that while different architectures fail differently, the underlying structure of failure remains consistent. This invariance across vendors rules out implementation flaws as the primary explanation. It points instead to a missing architectural primitive.
The proposed response is not to discard capability benchmarks, but to recognize their limits. Capability levels describe what systems can do. They do not describe whether systems can remain stable when it matters. Functional intelligence requires an explicit representation of significance that persists across context shifts and guides dominance when interpretations compete. The S-Vector is introduced as a control channel intended to fill that role, enabling generality without semantic collapse.
Seen together, these contributions form a closed loop rather than a series of disconnected critiques. The functional framing challenged mastery-based definitions of AGI. Observed failures across deployments confirmed the relevance of that challenge. The significance deficit identified the missing variable. Fracture–repair dynamics explained the behavior. The architectural proposal addressed the gap. AI Conversation Phenomenology emerges as the field required to study these dynamics systematically.
This work does not claim that current AI systems do not represent enormous capability evolution. It claims that progress has been measured along the wrong axis. Capability is not the same as stability. Functional AGI will not be defined by how many tasks a system can perform, but by how rarely meaning collapses when stakes are real. Closing that loop changes what intelligence looks like, how it should be evaluated, and what must exist for it to be trusted at scale.
