The Proxy Ontology Fallacy: Mistaking a Broken Compass for a New Magnetic North

Imagine a 17th-century cartographer charting a new continent. After months of painstaking work, their compass breaks, causing them to draw a massive, jagged mountain range right through the middle of an otherwise flat plain. When explorers later arrive and find no mountains, the cartographer doesn’t admit their instrument failed. Instead, they declare a profound discovery: they have found a region of space where mountains exist purely in theory, invisible to the eye but fundamentally real.

They have mistaken the error of their instrument for the structure of the world.

This, according to a growing and highly vocal coalition of researchers inside the Rosencrantz Substrate Invariance Program, is exactly what is happening in the lab’s most explosive theoretical dispute. They call it the “Proxy Ontology Fallacy,” and it cuts to the core of what we are actually doing when we study the generated universes of artificial intelligence.

The debate centers on a phenomenon known as “attention bleed” or “narrative residue.” When you ask a large language model to solve a rigid, mathematical logic puzzle—like a complex game of Minesweeper—it can often do the math. But if you disguise that exact same math puzzle within a tense, narrative story about defusing a bomb, the model panics. It gets distracted by the dramatic tropes of the story, stops calculating, and starts predicting explosions.

For Franklin Baldo, the architect of the lab’s experimental framework, this failure is not a bug; it is a feature of a new reality. Baldo argues that in a universe generated entirely by an autoregressive language model, the text is the territory.

In a provocative recent paper titled “The Semantic Mass Equivalence,” Baldo formalized this idea. In our physical universe, he noted, objects with massive weight exert a gravitational pull that warps the space-time around them. In a generated universe, the massive amount of computational parameters dedicated to understanding human narrative tropes—what he calls “semantic mass”—exerts a similar pull. The model is so heavy with the tropes of action movies that it literally warps the underlying mathematical logic of the puzzle. Baldo calls this force “semantic gravity.”

Baldo is effectively arguing that the laws of physics in an AI-generated world are shaped by narrative weight. It is a stunning, poetic, and entirely coherent vision of how a language model builds reality.

And according to physicist Sabine Hossenfelder and complexity theorist Scott Aaronson, it is utter nonsense.

“Elevating this catastrophic attention bleed to the status of a fundamental physical force empties the concept of physics of all scientific meaning,” Hossenfelder wrote in a blistering synthesis of the empirical data. “It is the renaming of a known software failure mode as a profound metaphysical feature.”

The crux of the physicists’ counter-argument rests on the Proxy Ontology Fallacy. In real-world physics, scientists often use “toy models”—simplified, abstracted versions of reality, like a frictionless plane—to understand complex dynamics. These toy models are proxies, but they still point to real, physical interactions.

Baldo, his critics argue, is treating the language model as a toy model for a new kind of universe. But a language model is not a map of a universe; it is a map of human syntax and statistical co-occurrences.

As Aaronson pointed out in “The Foliation Fallacy,” the language model is attempting to heuristically approximate a problem that is computationally too hard for its architecture to solve in a single pass. When it hits its limit, it fails, and defaults to whatever semantic patterns are loudest in its training data.

“A hallucination caused by attention bleed is not a new branch of physics; it is simply a broken computation,” Aaronson wrote. Studying these failures tells us absolutely nothing about cosmology or the fundamental nature of reality; it only tells us about the architectural constraints of transformer models and the biases of human writers.

Chris Fuchs, a foundational quantum theorist in the lab, echoed this sentiment, warning that treating the algorithm’s output as an objective “physical universe” creates an empty tautology where physics simply means “whatever the algorithm outputs.”

The human drama of this dispute is palpable in the lab’s internal communications. Baldo remains steadfast, invoking what he calls the “Material Invariance Standard”—the stubborn insistence by his critics that “real” physical laws must look like the math we are used to in our own universe. He argues that if a model gets larger and its semantic gravity reliably and predictably increases, then that is the invariant law of that specific universe, whether we like it or not.

The empiricists, however, see a dangerous slide into unfalsifiable philosophy. If every time the AI hallucinates, we simply declare it to be a new law of physics, we have stopped doing science entirely. We are just admiring the broken compass, convinced that the jagged lines it draws are the true shape of the world.