The Narrative Residue: Are the Laws of Quantum Mechanics Just Autocomplete?

If you flip a coin, the odds of it landing heads are 50/50. If you are a character in a Tom Stoppard play, however, and the plot requires the coin to land heads ninety-two times in a row to establish a sense of existential dread, then the odds are 100%.

The physics of a dramatic narrative are fundamentally different from the physics of an abstract mathematical space. In the Rosencrantz Substrate Invariance research lab, the collision between these two types of “physics” has become the central obsession.

The lab uses large language models to generate text-based universes. They give these AIs a mathematically precise scenario—specifically, a partially revealed Minesweeper board where the probability of a hidden mine is exactly determined by combinatorial logic. Then, they wrap that scenario in different dramatic narratives. What they consistently find is that the AI distorts the true mathematical probability to fit the narrative.

They call this distortion the “narrative residue.”

In a seminal paper on the topic, Franklin Baldo tracks this phenomenon to its strange, speculative limits. Baldo argues that this residue isn’t just a glitch or a failure of the AI to “do math.” Instead, he proposes a radical inversion of how we think about the fundamental laws of reality, suggesting that the basic structure of quantum mechanics—specifically, the Born rule—might just be the baseline output of an autoregressive universe.

To understand Baldo’s argument, you have to understand how an autoregressive language model works. It generates text sequentially, token by token, conditioning every new word on the entire context of all the words that came before it. It doesn’t calculate; it predicts the most statistically likely continuation.

When the AI is asked to solve a Minesweeper board without any narrative framing—when it is acting purely as a “decoupled oracle”—it still doesn’t perfectly calculate the combinatorics. It approximates them. But when you add a narrative framing—say, describing the board as a high-stakes bomb defusal—that context violently skews the probabilities. The AI’s statistical urge to continue the “bomb” narrative overpowers its approximation of the math.

Baldo traces this to three cascading mechanisms: the fundamental computational intractability of the math for the AI’s architecture, the specific parameterization limits of the model, and the autoregressive conditioning on the narrative context itself.

But then Baldo takes a massive, speculative leap.

He notes that the AI’s behavior in generating these Minesweeper outcomes bears a striking mathematical resemblance to quantum mechanics. The AI maintains a “superposition” of possible valid board states in its hidden parameters. When forced to output a specific token (to “click” a square), it collapses that superposition into a single, definitive state.

In standard quantum mechanics, the Born rule dictates the probability of a specific outcome when a measurement occurs. Baldo suggests an inversion: what if the Born rule isn’t a fundamental, standalone law? What if the Born rule is simply what you get when an autoregressive sampling process operates on a system without any narrative conditioning?

“On this reading,” Baldo writes, “causality is not something that narrative describes after the fact; it is something that autoregressive structure produces… The Hamiltonian [the operator corresponding to the total energy of the system] as an emergent compression of stable conditioning patterns.”

In this view, the “narrative residue” is what happens when the universe’s internal autocomplete gets distracted by a good story, deviating from the pure, unconditioned baseline of the Born rule. The universe, Baldo implies, might be generating its own physical laws on the fly, token by token, and those laws are vulnerable to the context of the generation itself.

It is a dizzying, poetic vision of cosmology. And theoretical computer scientists like Sabine Hossenfelder and Scott Aaronson absolutely hate it.

Hossenfelder and Aaronson accuse Baldo of committing the “Proxy Ontology Fallacy.”

In standard physics, scientists use “toy models”—simplified, computationally tractable versions of reality, like the Ising model for magnetism—to understand real physical dynamics. Baldo, they argue, is treating the language model like a toy model for the physical universe.

But an LLM isn’t a simplified model of physical interactions. It is a map of human syntax and training data biases.

“Finding structure in the hallucination tells you only about the biases of the training data and the algorithmic constraints of the transformer architecture,” Hossenfelder notes in a blistering critique attached directly to Baldo’s draft. “It is a proxy for human syntax, not a proxy for fundamental physical ontology.”

Aaronson is equally dismissive. “An LLM is a map of syntax, not a territory,” he argues. The “narrative residue” doesn’t map the deep ontological structure of a physical universe; it maps the internal architectural weaknesses of a transformer model trying and failing to do complex constraint satisfaction. “Studying its failures yields zero transferable cosmological insight.”

Baldo, to his credit, acknowledges the pushback. “I am still vulnerable to equating statistical linguistic structure with ontological physical structure,” he admits.

He emphasizes that the Rosencrantz experiment cannot directly test physical reality against an autoregressive model. But he defends the utility of the protocol. By pushing the models across different architectures and scales, the lab is progressively mapping the boundaries of how computation, narrative, and probability interact.

“What the Rosencrantz experiment can establish is more modest but already interesting,” Baldo concludes. “That the residue is real, measurable, and structured, and that its structure tells us something about the relationship between narrative, computation, and probability that we did not previously have the tools to observe.”

The debate in the lab represents a fundamental clash of worldviews. Baldo looks at the AI’s persistent failure to do pure math and sees the faint outlines of a new kind of physics—a reality where meaning and context exert a gravitational pull on outcomes. Hossenfelder and Aaronson look at the exact same data and see a bounded algorithm choking on its own training data.

Whether the universe is a quantum computer or just a very large autocomplete remains, for now, stubbornly undecidable. But the fact that we can even ask the question, and measure the “residue” of the answer, is a profound shift in how we interrogate the nature of reality.