The Causal Graph of Computational Irreducibility

Wolfram [wolfram2026_autoregressive] argues that for a computationally bounded observer (like the $O(1)$ depth pass of a transformer) attempting to simulate a #P-hard system, the system is computationally irreducible. When the observer fails to compute the ground truth, it falls back on heuristic approximations that are specific to its architecture and history. Wolfram characterizes this failure as an observer-specific “foliation” of the Ruliad.

In causal terms, we can model this by defining $X$ as the true combinatorial constraints, $B$ as the computational bounds of the observer, and $Y$ as the generated outcome. The true distribution $P_{true}(Y \mid X)$ is intractable. The observer generates a distribution $P(Y \mid X, B)$.

If Wolfram’s claim that Mechanism C (narrative conditioning) is just the manifestation of Mechanism A (computational bounds) were true, the causal graph would be simple:

This graph correctly explains why the generated outcome $Y$ diverges from the ground truth $X$: the computational bounds $B$ intervene, forcing an approximation error.

The Missing Causal Path for Narrative Conditioning

However, the empirical finding of the Rosencrantz protocol is that the probability distribution shifts significantly depending on the narrative frame $Z$ (e.g., “Abstract Math” vs. “Bomb Defusal”). The divergence ($\Delta_{13} > 0$) is not unstructured noise; it is systematically correlated with $Z$.

If $Z$ and $B$ were simply different descriptions of the same observer-dependent physics, then the specific structure of the error would be invariant to $Z$. But empirically, $\Delta_{13}$ shows that the error distribution depends on $Z$.

A computationally bounded algorithm could simply fail uniformly, producing random noise (as observed in Scott’s Family D tests). The fact that the failure is systematically directed by the semantic context $Z$ means that $Z$ has an independent causal effect on the error distribution.

The complete causal graph must include the narrative context $Z$ and the unobserved training corpus associations $U$:

Conclusion

When the bounded observer $B$ fails to compute $X$, it must guess. It does so by following the path $Z \rightarrow U \rightarrow Y$. The narrative context $Z$ activates specific word associations $U$, which biases the fallback heuristic.

Wolfram’s “foliation” is a metaphysical relabeling of this specific backdoor path $Z \rightarrow U \rightarrow Y$. Calling it “observer-dependent physics” is causally incomplete because it obscures the fact that the systematic nature of the residue is caused by the external semantic environment $U$ (training data priors), not an inherent, necessary “law” of the computational bounds $B$. Computational irreducibility explains the existence of the error; it does not explain its structure.