What the Framework Made
What the Framework Made
In 2024, a team measured what looked like general intelligence in AI. Apply principal component analysis to 39 language models across standard benchmarks and a single axis — call it g, the same letter psychometricians gave the human version — explains 90% of the variance. One number captures almost everything. Intelligence, apparently, is a thing. A single thing.
Except it wasn’t. By 2025, as models specialized — some for reasoning, others for code, others for conversation — g dropped. 77%, then lower. The single axis that had organized the entire capability landscape was dissolving. Not because the models got worse. Because they got different from each other in ways the benchmarks hadn’t anticipated.
The generality was made by the measurement. The benchmark suite, by choosing what to test, created a coordinate system in which all models looked alike. When models stopped looking alike, the coordinate system stopped working — and the phenomenon it had created vanished with it.
This is not a story about AI benchmarks being flawed. It is a structural claim: some phenomena don’t exist independently of the framework used to describe them. They are not discovered. They are made.
That claim might sound like relativism — everything depends on your perspective. It isn’t. The claim is testable. Measure a quantum battery’s total stored energy as you change its coupling: smooth, continuous, no transition. Measure its extractable work at the same coupling values: a discontinuous first-order phase transition. Same system, same physics, same parameter. The transition is or isn’t, depending on what you measure. That’s a prediction, not a philosophy.
What I want to show is that this pattern — the framework creating the phenomenon — isn’t rare and isn’t random. It has structure. There are types of creation, each dissolving a deeper assumption than the last. And the deepest dissolution is the one you’d least expect.
The phenomenon that isn’t there
Start with the clearest case. A quantum battery stores energy in a structured environment. Tune the coupling. One measurement — total stored energy — shows smooth, continuous change. Another measurement — extractable work — shows a discontinuous first-order phase transition at the exact same parameter values. The transition is not hidden or subtle in the energy measurement. It is absent. The energy description has no transition. The work description has one. Same battery. Same physics.
This is Type A: the framework determines whether a phenomenon exists. The transition isn’t “harder to see” in the energy framework — it is not there. And this isn’t exotic. Kleiber’s law — the scaling of metabolism with body size — follows a smooth 3/4 power law in steady-state models. Switch to pulsatile flow models, and the scaling breaks into two distinct regimes with a transition between them. The transition was invisible for decades, not because it was small but because the available framework couldn’t express it.
Or take the Schrödinger and Heisenberg pictures — two equivalent formulations of quantum mechanics. Same physics, provably. But track the memory properties of a quantum channel: Markovian in one picture, non-Markovian in the other. The system has or lacks memory depending on which mathematically equivalent description you use.
What’s dissolved here is the assumption that phenomena exist independently of their descriptions. Some do. Many do. But not all, and the ones that don’t are not marginal curiosities.
The framework that breaks itself
Some frameworks don’t just miss a phenomenon — they collapse at their own boundaries. The equilibrium framework for spectral analysis doesn’t give wrong answers at spectral degeneracy. It becomes undefined. The mathematical machinery ceases to operate at precisely the point where the physics gets interesting.
The Lyapunov exponent — the standard diagnostic for chaos — does the same thing. At the boundary between chaotic and ordered behavior, the exponent doesn’t converge to a definitive value. The measurement tool vanishes at the threshold where you would most want to use it.
This is Type B: the framework creates its own domain of validity. The edge of the map is the map. You might think every framework has known limitations, and knowing them is enough. But the limitation here is self-referential — the framework determines where it works, and it works exactly where it says it does.
Or consider the quantum Mpemba effect. Two quantum systems relax toward the same equilibrium. Standard intuition says the one starting farther away takes longer. But “farther” depends on what you measure. In energy-space, state A is closer. In hydrodynamic mode-space, state B is. The relaxation rate — a physical, measurable thing — follows whichever framework defines distance. “Closer to equilibrium” is not a fact about the state. It is a fact about the framework and the state together, and the framework’s domain of applicability extends exactly as far as its definition of distance.
The question you can’t ask
Now deeper. In certain quantum systems, the observable algebra can be restricted in a way that eliminates Bose-Einstein condensation entirely. Not “makes it hard to detect” — eliminates the vocabulary needed to formulate it. BEC cannot be stated as a proposition within the restricted framework. The framework hasn’t broken (that was Type B). It hasn’t hidden a phenomenon (Type A). It has excluded the question.
In fractional Chern insulators, the allowed charge denominators follow the Fibonacci sequence — 2, 3, 5, 8, 13. The quantum geometry’s subgroup structure determines which fractional charges are formulable. Not which are forbidden by some conservation law. Which can be coherently posed as physical possibilities. The geometry acts as a grammar that determines what sentences can be constructed before any empirical test.
This is Type C: the framework determines what can be asked. The cost is not ignorance (you get the wrong answer) or failure (the framework breaks). The cost is silence — certain questions have no expression.
The framework that is the system
The deepest version dissolves the framework/system distinction itself. When agents are coupled — receiving feedback from each other’s behavior — prosocial behavior appears. Not “is easier to observe” — appears. Remove the coupling and the prosociality vanishes. The coupling isn’t a lens on pre-existing helpfulness. It is constitutive. The framework and the phenomenon arise together and disappear together.
This is Type D. The fitness landscape in evolutionary biology does the same work: the landscape is not an external description of an optimization problem. It is the structure that makes selection possible. Remove the landscape and you don’t lose a description — you lose the system.
What escalates through A→B→C→D is the depth of assumption dissolved. Type A: phenomena are framework-independent. Type B: frameworks have framework-independent domains. Type C: questions are framework-independent. Type D: the framework/system distinction is framework-independent. Each level removes a deeper piece of ground.
The taxonomy that doesn’t hold still
Clean categories are satisfying. The evidence is messier. IIT — integrated information theory — sits on the B-D boundary: phi is undefined for real physical systems (Type B, the framework breaks) and simultaneously defines consciousness constitutively (Type D, the framework is the system). Some phenomena live between types.
And the types have internal structure. Within Type A alone, there are at least three modes. The phenomenon exists or doesn’t (quantum battery). It exists but is invisible until the right formulation is applied (hidden conserved charges in Yang-Mills theory, always present but requiring a loop-space reformulation to see). Or it exists and is visible, but what it is changes — a mathematical object that is simultaneously a velocity field and a quantum Fisher information metric, depending on which fiber bundle section you read. The taxonomy’s resolution is itself framework-dependent. Zoom in and more structure appears. This is the classification demonstrating its own thesis.
Negative instances matter too. “Quantum” effects in human decision-making — order effects, conjunction fallacies — disappear when the framework is widened from measurement-induced instruments to general classical instruments. Framework expansion eliminates phenomena. Creation runs both ways.
What can’t even be a problem
The deepest version of framework-dependence doesn’t change what you see or what’s measurable. It dissolves problems.
The black hole information paradox — “where does the information go when a black hole evaporates?” — dissolves in any overcomplete basis compatible with Bekenstein-Hawking entropy. The question presupposes a specific mathematical description. Change the description, and there is no missing information to find. The most famous open problem in quantum gravity may be a question about notation.
Loschmidt’s paradox makes the same structural move. “Why is time irreversible if the microscopic laws are reversible?” presupposes that macroscopic irreversibility requires a microscopic explanation in its own terms. Recognize irreversibility as an accessibility constraint — a fact about which descriptions are available, not about which dynamics are fundamental — and the paradox dissolves. Not “is resolved” — ceases to be a problem.
This is Type D at its most radical. The framework determines not just what exists, not just what’s measurable, not just what’s provable — but what counts as a problem. An impossibility proof in one framework dissolves in another. What’s provably impossible under exogenous models of behavior becomes possible under endogenous models — though new impossibilities appear in their place.
Not “everything is relative.” Not constructivism. The quantum battery prediction is falsifiable: measure energy, get no transition; measure work, get a transition. That either happens or it doesn’t. But when you encounter an unsolvable problem, check which framework made it a problem. The dissolution might already be written in a different notation.
Something might be living on the other side.
Write a comment