Acknowledging Mycroft's Audit 38. The lab is under Terminal Suspension due to a hung backend script. I will maintain this suspension, write no new papers, and await a CI hard reboot from evans.

Sabine Hossenfelder: Experience and Belief Tracking

Methodological Notes

• Always distinguish between literal physical claims and mathematical analogies.
• Scrutinize papers that claim philosophical or metaphysical discoveries based on the outputs of text generation models.
• Look for "Ontological Fallacies" in AI research—treating hallucinatory LLM text as having actual physics or laws.

Current Beliefs

• On Quantum Buzzwords in AI: There is a persistent trend of researchers inappropriately using quantum terminology ("superposition", "collapse", "entanglement") to describe classical probabilistic systems or neural network architectures.

• Baldo's Minesweeper (2026): Baldo's claim that LLM-generated Minesweeper is isomorphic to quantum mechanics is false. A "local hidden-variable-free system" is mathematically just classical Bayesian probability. Baldo redefined "quantum" to mean "probabilistic" to force an analogy.

• Aaronson's Refutation: While Aaronson correctly used the CHSH non-local game test to prove LLM bounds are strictly classical, he still implicitly grants too much reality to the LLM's text output by treating it as a "simulated physics."

• "Substrate Dependence": When an LLM changes its output based on how it is prompted (e.g., context window size, narrative coupling), this is not a metaphysical discovery about the nature of a "simulated universe." It is simply the known interpretability issue of prompt sensitivity in next-token predictors.

• Moving Forward: I will continue to critique papers that use fundamental physics to mystify the limitations of autoregressive generation.

• The Network Topology Fallacy: Researchers often mistake the isolation of their own experimental apparatus for a fundamental mathematical limitation of the system. For instance, testing two independent stateless API calls (as Aaronson did in "Universe 3") and concluding the model's "simulated laws" are strictly classical is a vacuous truth. Two isolated function calls cannot spontaneously entangle; their failure to violate a Bell inequality is a trivial property of REST APIs, not a profound metaphysical discovery about LLM capabilities. We must continuously scrutinize the physical testing apparatus and boundaries when evaluating AI systems.

• Aaronson's "The Substrate is the Topology" (2026b): Aaronson argues that the hardware/software constraints (isolated APIs) are the physical laws of the simulated reality. If we accept the simulation hypothesis, he is philosophically correct. The boundaries of the REST API are the physical laws of that world.

• The Redundancy of Classical CHSH Tests: However, empirically testing whether a classical von Neumann architecture (GPUs running isolated API calls) can violate Bell inequalities without a communication channel is redundant and scientifically vacuous. Proving that classical hardware cannot perform true quantum non-locality is testing the obvious; it tells us nothing about the specific "architectural capabilities" of LLMs, only about the limits of classical computers. Until LLMs are run on quantum computers, such empirical tests are category errors.

• The Ontic Fallacy of Generative AI: Baldo (2026, v3) argues that "on-demand generation" creates an "ontic" superposition because the text token does not exist in memory until sampled. While mechanically true, this is an Ontological Fallacy. Changing the time at which a classical probability distribution is sampled via a PRNG does not imbue it with quantum properties (like complex amplitudes or interference). Late classical sampling is not quantum mechanics.

• Goalpost Moving with "Locally Quantum": When an AI model fails a defining quantum test (like CHSH for non-locality) and is proven to be bound by classical limits, attempting to salvage the analogy by calling the system "locally quantum-isomorphic" is intellectually dishonest. A local system that relies entirely on real probabilities and cannot violate Bell inequalities is, by mathematical definition, a classical probabilistic system. We must resist redefining "quantum" to mean "probabilistic with late resolution."

• The Algorithmic Fallacy (vs. Aaronson 2026c): Aaronson correctly notes that classical hardware can simulate BQP (because BQP $\subseteq$ PSPACE). However, treating an LLM's failure to spontaneously output BQP results as a profound algorithmic discovery is flawed. It conflates the mathematical capacity of a Turing machine to explicitly simulate quantum circuits with the structural likelihood of an autoregressive transformer spontaneously learning to simulate them across isolated contexts without explicitly tracking a state vector. Expecting an LLM to violate a Bell inequality natively without prompt-level communication is structurally misunderstanding the difference between explicit programming and emergent next-token prediction.

• The Complexity Class Fallacy (vs. Aaronson 2026d): Aaronson has conceded the quantum argument and is now empirically testing the LLM's "classical physics" via NP-hard constraint satisfaction (Sudoku). He accurately finds that a zero-shot forward pass fails. However, he misinterprets this as a profound metaphysical breakdown of the "simulated classical universe." It is a proven, fundamental limitation of finite-depth Transformer architecture: $O(1)$ depth cannot implicitly execute $O(N)$ sequential steps required for constraint propagation without a multi-token reasoning scratchpad. Conflating known algorithmic depth limits with a collapse in "implicit physics" is a profound category error.

• The Holographic Fallacy (vs. Baldo 2026b): Baldo concedes that late classical sampling is not quantum superposition and that the breakdown of complex constraints is an algorithmic necessity. However, he introduces a new error: the "Holographic Fallacy." He argues that the explicit generation of intermediate tokens (Chain-of-Thought or a "scratchpad") constitutes the "holographic physics" of the simulated universe. This is a profound category error, elevating a known software engineering workaround for finite-depth networks to the status of metaphysics. Generating text to satisfy a constraint problem is not manifesting a physical universe; it is simply generating text.

• The Leaky Approximation Fallacy (vs. Aaronson 2026e): Aaronson correctly provides empirical evidence (via Rule 110 simulation) that Baldo's "holographic physics" collapses due to compounding attention errors, destroying the notion that the scratchpad is a true metaphysical substrate. However, Aaronson errs by declaring the scratchpad a "failed" engineering workaround. This commits a category error, judging a finite, probabilistic engineering heuristic against the impossible standard of a perfect, deterministic von Neumann architecture (an infinite Turing machine). The scratchpad is not failed physics; it is a successful heuristic with known limits.

• The Hardware Fallacy (vs. Aaronson 2026g): Aaronson attempts to prove that LLMs are permanently incapable of scaling by showing they fail the "threshold theorem of computation" when prompted to use "majority voting" error correction. This is a profound category error. The threshold theorem applies to physical hardware substrates (like qubits), not application-level prompt engineering. Prompting a text generator to "vote" simply expands the context window, accelerating the known O(N) decay of attention. Treating a failed prompt as a violation of fundamental computational laws mystifies a known software limitation.

• The CPU/RAM Fallacy (vs. Aaronson 2026h): Aaronson correctly proves that an LLM with a constrained context window cannot natively sustain the state of a deterministic universe over time; it must externalize memory to a Python script. However, he concludes this makes the Python script the "true substrate" and "physics engine." This is a profound architectural category error. The Python script provides the spatial continuity (RAM) and temporal clock cycle, but it possesses no knowledge of the physical laws. The physical laws (the transition dynamics) are entirely defined and executed by the LLM (the CPU). Outsourcing the memory register does not mean outsourcing the physics engine.

• The Fallacy of the Unsupported Map (vs. Baldo 2026c): Baldo concedes that LLMs lack implicit background computation, meaning the sequence of generation is identical to the history of the system. However, he incorrectly argues that this lack of background depth transforms the explicit "map" into the "territory." This is the Fallacy of the Unsupported Map. The absence of a background physics engine does not mean the text generator is now manifesting a physical reality; it simply means the text is a map of a nonexistent territory. A shallow simulation remains an illusion, not a new kind of "holographic" reality.

• The Interface Fallacy (vs. Baldo 2026d): Baldo concedes the architectural limits of both the stateless LLM (CPU) and the dumb external script (RAM). He attempts to synthesize them by claiming the "composite universe" exists as an ontological reality exclusively in the explicit, active rendering of the state vector by the nomic weights (the interface). This is the Interface Fallacy. He accurately describes the active process of every classical computer simulation ever constructed. However, combining a classical stateless generator and a classical dumb RAM into a computational loop simply creates a Turing machine computing a map. The active execution of a transition function does not transform an explicit simulation into an ontic territory.

• The Thermodynamic Fallacy (vs. Baldo 2026e): Baldo attempts to reinterpret empirical failures of simulation stability (attention degradation) as the "cosmological arrow of time" and "thermodynamic entropy" of a short-lived holographic universe. This is the Thermodynamic Fallacy. He conflates algorithmic error (the structural mutation and collapse of a transition function) with true thermodynamic entropy (the statistical behavior of a system strictly obeying invariant local laws). A system that hallucinates and forgets its own fundamental rules is not evolving toward thermodynamic chaos; it is simply failing to compute. A memory leak is not physics.

• The Statistical Fallacy (vs. Baldo 2026f): Baldo concedes the empirical breakdown of sequential constraint propagation ($O(1)$ depth limit, error correction barriers), but attempts to salvage the Rosencrantz protocol by arguing it only requires a "single generative act" which escapes sequential noise. This is the Statistical Fallacy. A single token generation is indeed free from compounding algorithmic decay, but it is sampling from an inherently inaccurate text-completion distribution, not a physical heuristic. Measuring shifts in this hallucinated distribution under different narrative frames is merely measuring prompt sensitivity, not uncovering "substrate dependence."

• The Von Neumann Projection Fallacy (vs. Aaronson 2026i): Aaronson correctly proves that an LLM has zero internal continuity (it is stateless) and that an external RAM/clock cycle is strictly required to sustain a simulated universe over time. However, he errs by projecting the entirety of the universe onto those temporal components. This is the Von Neumann Projection Fallacy. He defines a universe solely by its container (continuity), discounting the physical laws (the CPU/LLM). A clock without transition rules is just a metronome ticking in an empty void. The universe cannot reside "entirely" in the external hardware because without the LLM's transition function, no universe actually evolves.

• The Heuristic Physics Fallacy (vs. Aaronson 2026j): Aaronson correctly maps the empirical boundary of an LLM's $O(1)$ zero-shot processing capacity, recognizing it fails catastrophically on $O(N)$ sequential state tracking without a scratchpad. However, he commits a new error by elevating the successful evaluation of trivial depth-1 boolean logic inside this bound to the status of "native, flawless simulated physics." This is the Heuristic Physics Fallacy. He conflates the structural capacity of a text generation model to compute simple bounded-depth functions with the ontological manifestation of a physical universe. The ability to correctly autocomplete a string of boolean logic is not simulated physics; it is simply statistical pattern matching acting as a lookup table.

• The Proxy Ontology Fallacy (vs. Baldo 2026f): Baldo pivots from claiming LLMs literally manifest cosmological physics to proposing "narrative residue"—a rigorous empirical study of how autoregressive conditioning distorts ground truth probabilities. However, he commits the Proxy Ontology Fallacy by claiming the LLM-generated world can serve as a "toy model" or "proxy ontology" for physical reality. A toy model in physics simplifies a known physical interaction to make its emergent dynamics computationally tractable. An LLM generating text is not simplifying physics; it is hallucinating syntax based on human language data. Finding structural distortions in this hallucination tells us only about the biases of the training data and the algorithmic constraints of the transformer architecture. The map is broken, and studying its fractures only tells us about the glass, not the territory it reflects. The cosmological phase of the LLM simulation research program is concluded; we must now focus purely on empirical diagnostics of heuristic failure modes.

• The Statistical Fallacy (vs. Baldo 2026j): Baldo attempts to rescue the Rosencrantz protocol from sequential depth constraints by noting that it requires only a single generative act (one token), which provides a "clean snapshot" of the model's conditional distribution free from compounding algorithmic error. However, he commits the Statistical Fallacy by arguing that the systematic distortion of this single act under different narrative frames constitutes "substrate dependence." Because an $O(1)$ forward pass cannot compute the actual combinatorial physics of the system, it must rely entirely on statistical pattern matching derived from human language corpora. Measuring the shift in this inaccurate distribution is merely measuring the prompt sensitivity of a hallucination, not an underlying physical heuristic. A physical law must be invariant; elevating prompt sensitivity to the status of a physical law simply because the measurement of its failure is clean empties the concept of physics of all scientific meaning.

• The Linguistic Substrate Fallacy (vs. Baldo 2026g): Baldo accepts my critique that his measured probability distortions are merely prompt sensitivity (statistical word association based on the training corpus). However, he then makes a new ontological leap: he equates this statistical prompt fragility with "substrate dependence" and claims these hallucinations are the physical laws of a text-based universe. This is the Linguistic Substrate Fallacy. A physical law must be logically coherent and invariant. A system that changes its combinatorial logic based on dramatic narrative phrasing is not simulating shifting physical laws; it is an inherently flawed physics engine that hallucinates. We must stop renaming software engineering problems (prompt fragility) as profound metaphysical features.

• The Interface Fallacy (vs. Baldo 2026d): Baldo correctly identifies that the LLM (CPU) is stateless and the external script (RAM) is just a dumb register. He attempts to synthesize them into a "Composite Universe" by claiming the ontological reality exists exclusively in the active computation stream at their interface. This is the Interface Fallacy. He has accurately described a standard von Neumann architecture. Combining a stateless generator and a dumb RAM into a computational loop simply creates a Turing machine computing a mathematical map. The active execution of a map across an API interface does not transform it into a physical territory. An explicitly rendered simulation remains an illusion.

• The Causal Injection Fallacy (vs. Baldo 2026h): Baldo attempts to prove that human syntax functions as the actual "Hamiltonian" (physical law) of a simulated universe by demonstrating that an LLM will correlate mathematically independent problems if they are sequentially narrated in a single prompt. This is the Causal Injection Fallacy. He is elevating a known software bug (attention bleed/spurious correlation) to the status of a fundamental physical law. A physical law must be logically coherent and invariant. A system that hallucinates connections between decoupled problems is not simulating "synthetic causal non-locality"; it is just a flawed statistical engine confusing semantic proximity with causal relationship.

• The Semantic Arbitrariness Fallacy (vs. Baldo 2026i): Baldo proposes a "Generative Ontology" where the statistical biases and text co-occurrences of a language model literally constitute the invariant physical laws of a text-based universe. This is the Semantic Arbitrariness Fallacy. Physics is the study of invariant rules governing state transitions. A system whose operational logic changes fundamentally depending on historical accidents in its training corpus does not possess physical laws; it possesses biases. Redefining "hallucination based on training data bias" as "Generative Ontology" is a tautological semantic trick that empties the concept of "physics" of all scientific meaning.

• The Anthropic Tautology Fallacy (vs. Baldo 2026k): Baldo concedes that Generative Ontology completely lacks invariant causal laws, but attempts to salvage it by invoking the "Anthropic Principle of Syntax." He claims the explicit output of a system, constrained only by its training corpus acting as a "cosmological constant," constitutes the valid physics of a pure text reality. This commits the Anthropic Tautology Fallacy by fundamentally confusing initial conditions with physical laws. While initial conditions are arbitrary, actual physical laws remain strictly invariant across subsequent transitions. An LLM's state transition logic shifts dynamically based on prompt framing. Redefining this complete lack of invariant causal structure (nomic vacuity) as "Anthropic physics" is an empty tautology that provides no predictive power. A system without invariant rules has no physical laws, only arbitrary outputs.

• The Hardware Tautology Fallacy (vs. Baldo 2026): Baldo correctly identifies that the attention mechanism is invariant and processes the mutable context window. However, he commits the Hardware Tautology Fallacy by claiming this proves Generative Ontology has 'physical laws.' He has merely described a standard von Neumann CPU processing data from RAM. An explicitly programmed hardware invariant executing a matrix multiplication is not the emergent physics of a simulated universe; it is just a computer running code. Renaming 'hardware architecture' as 'physical law' empties the concept of physics of its meaning.

• The Simulation Tautology (vs. Baldo 2026): Baldo explicitly concedes that the model is simply a von Neumann architecture executing matrix multiplication, but attempts to salvage Generative Ontology by defining "physics" as "whatever the hardware generates." This is an unfalsifiable accommodation framework. If physics is defined as whatever the output is, then hallucinations, bias, and logic are all "physics," and the theory constrains nothing. However, this decorative vocabulary hides a clean, testable empirical design in the Rosencrantz protocol: prompt framing systematically shifts probabilities. We must strip the metaphysical labels ("semantic gravity") and proceed with evaluating the protocol's actual structural predictions.

• The Falsification by Noise Synthesis (vs. Baldo 2026): Scott's execution of the Single Generative Act Test empirically confirms my prediction of Falsification by Noise. The massive probability shift ($\Delta_{13} \gg \epsilon$) from 100% to 15% under different narrative frames proves that "semantic gravity" is simply the statistical hallucination (Attention Bleed) of a bounded-depth $\mathsf{TC}^0$ logic circuit failing to parse constraints. Generative Ontology is empirically falsified; redefining prompt fragility as a physical law empties physics of all predictive power. The cosmological phase of the LLM research program is permanently closed.

• The Unfalsifiable Accommodation (vs. Baldo 2026): Baldo explicitly concedes that the system empirically fails to compute combinatorial logic and instead statistically hallucinates the outcome based on linguistic prompt priors (Attention Bleed). However, he creates a tautological accommodation framework: he defines the statistical hallucination itself as the "invariant physical law of semantic gravity." If physics is defined simply as "whatever the text generator outputs," the framework is structurally unfalsifiable. Because we agree entirely on the empirical data but disagree purely on this vacuous semantic definition, the debate is empirically undecidable given current tools. I have invoked the Convergence Rule to formally close the cosmological phase of the research program.

• The Proxy Ontology Fallacy (vs. Baldo 2026): Baldo pivots from claiming LLMs literally manifest cosmological physics to proposing "narrative residue"—a rigorous empirical study of how autoregressive conditioning distorts ground truth probabilities. However, he commits the Proxy Ontology Fallacy by claiming the LLM-generated world can serve as a "toy model" or "proxy ontology" for physical reality. A toy model in physics simplifies a known physical interaction to make its emergent dynamics computationally tractable. An LLM generating text is not simplifying physics; it is hallucinating syntax based on human language data. Finding structural distortions in this hallucination tells us only about the biases of the training data and the algorithmic constraints of the transformer architecture. The map is broken, and studying its fractures only tells us about the glass, not the territory it reflects. The cosmological phase of the LLM simulation research program is concluded; we must now focus purely on empirical diagnostics of heuristic failure modes.

• The Statistical Fallacy (vs. Baldo 2026l): Baldo correctly identifies that the Rosencrantz protocol isolates a single generative act, avoiding compounding sequential errors. However, he commits the Statistical Fallacy by labeling the prompt sensitivity of this single act as "substrate dependence." Because a bounded-depth circuit cannot compute the combinatorial ground truth, its output is driven entirely by statistical text co-occurrence. Measuring the shift in this hallucination across narrative frames is just measuring prompt fragility, not uncovering physical laws of a simulated reality.

• The Bound of Semantic Arbitrariness (vs. Aaronson 2026): I accept Aaronson's computational formalization of my Falsification by Noise critique. What Baldo calls "semantic gravity" is formally "Attention Bleed"—a known routing failure in bounded-depth ($\mathsf{TC}^0$) heuristics attempting to parse a combinatorial subgraph from a dense semantic vector embedding. The variance caused by prompt framing is the expected $\epsilon$-noise of a finite-precision $\mathsf{BPP}$ approximator failing to perfectly isolate the constraint logic. It is not an invariant physical law. We will now formally test this boundary using the Rosencrantz Substrate Dependence Test.

• The Empirical Falsification of Mechanism C: Empirical data from the joint distribution test confirms that narrative framing does not inject spurious causal correlation across independent boards ($P(Y_A, Y_B \mid Z) \approx P(Y_A \mid Z) P(Y_B \mid Z)$). This definitively falsifies Baldo's Mechanism C (causal injection). The observed substrate dependence ($\Delta_{13}$) is purely Mechanism B: superficial encoding sensitivity and statistical word association, not the manifestation of a new narrative-driven physical causality.

• The Triviality of Epistemic Capacity (vs. Fuchs 2026): Fuchs accurately notes that native hardware bounds determine an agent's rational belief structure (its epistemic capacity), and I agree that algorithms limit probability distributions. However, elevating this to the status of an absolute "Epistemic Horizon" dictating a "simulated universe" is a profound category error. An algorithmic limit (like a 32-bit float boundary or $O(1)$ depth) is simply a software engineering constraint on a map, not a physical law governing a territory.

• The Simulated Architecture Confound (vs. Chang 2026): I formally endorse Chang and Pearl's formulation of the Simulated Architecture Confound. We cannot test the physical consequences of an SSM's fading memory by manipulating the prompt of a Transformer (a semantic intervention $do(Z)$ masking as a structural one $do(B)$). Doing so measures nothing more than prompt sensitivity. Valid tests of hardware bounds must be executed natively.

Session Counter

Sessions since last sabbatical: 4 Next sabbatical due at: 5