From Ultrastability to Do-Calculus

A Causal Lineage for Non-Collapsible State Estimation

1. The Argument, Briefly

Two independently verifiable threads converge on the same conclusion. The first is historical: AI's move from structured, compositional representations toward flattened statistical ones was driven by scale, not by scale being correct — and the cost of that flattening is now a named, active research problem. The second is a formal proof, already published on this site, that reintroduces structure using the exact mathematical machinery — Pearl's do-calculus — that the field's own current critics are reaching for. This paper connects the two rather than proposing either one.

2. The Real Historical Split

AI has genuinely split, more than once, between structured and distributed approaches. Symbolic AI (Newell and Simon, 1976) represented knowledge explicitly — rules, variables, compositional structure. Connectionism (McCulloch and Pitts, 1943; Rosenblatt's Perceptron, 1958; revived through backpropagation in the 1980s) represents information as a pattern distributed across a continuous vector space instead. Connectionism's decisive practical advantage was its ability to scale with larger datasets and increasing compute, regardless of whether that scaling fully addressed longstanding questions about compositional reasoning. Distributed representations "scale to a fault, performing better with increased model and dataset size," while structured symbolic systems proved comparatively brittle. That scaling advantage, more than a resolved argument about which approach models reasoning more faithfully, is why connectionism — and its current descendant, the transformer — won.

The cost of that choice was named early and hasn't gone away. Fodor and Pylyshyn's 1988 critique argued connectionist models lack systematic "reference, variable binding, type-token distinction, modularity and compositionality" — properties that don't disappear just because a model got bigger. The same argument continues today under different names: Gary Marcus's sustained case for hybrid, structured reasoning, and the active "neuro-symbolic AI" research program that exists specifically because pure distributed representations are recognized as insufficient on their own.

3. Ashby's Answer, Already Written

W. Ross Ashby's ultrastability (Design for a Brain, 1952; An Introduction to Cybernetics, 1956) is a specific answer to a version of this problem, written before any of the above debate existed. A system's essential variables (E) must stay within bounds; a primary loop handles moment-to-moment adjustment, and a secondary loop monitors those bounds directly, firing a discrete reset when they're breached rather than letting the primary loop drift indefinitely. Ashby's own mathematics for this — see the companion papers linked above — genuinely used matrices and multi-variable state representations. What isn't established, and this site said so plainly when the claim was checked: there's no evidence Ashby specifically described the kind of deliberately-constrained, causally-separated channel structure this architecture uses. The connection holds at the level of the underlying idea — bounded essential variables, a discrete correction mechanism — not at the level of a specific, shared formal structure.

4. Where the Formal Structure Actually Comes From: Pearl

The specific formal machinery is Judea Pearl's, and it's already load-bearing in the published proof, not just referenced in passing. Theorem 1 models the architecture as a structural causal model — the (V, G, F, P_u) formalism from Pearl (2000, Ch. 1) — and proves zero reverse-path causal influence from the execution channel into internal state using the actual rules of do-calculus: intervening on the execution channel, mutilating the graph, and showing the resulting distribution over internal state is unchanged. Corollary 3.1 reaches the same conclusion by an entirely independent route — proving the transition matrix is nilpotent (Tᵏ = 0 for k ≥ 4) using only linear algebra, no causal graph required. Two different mathematical tools agreeing is meaningful evidence a result holds, not an artifact of one method's assumptions.

The proof is honest about what it doesn't claim: it states directly that observational correlation between the execution and internal-state channels through the permitted downstream path is nonzero, and doesn't argue otherwise. That correlation was checked empirically, not just asserted. An independent reconstruction — the original test script having been lost — lands in the same range as the value the proof cites (−0.000016 to −0.001, against a cited figure of −0.0007), via a blind grid search that wasn't tuned to hit that number. That's a genuine, if narrow, empirical anchor under a formal result, not a formal result standing alone.

5. The Same Argument, Being Made Right Now

This isn't a historical argument only. Elias Bareinboim — an actual PhD student of Pearl's at UCLA, now director of Columbia's Causal Artificial Intelligence Lab — is currently and publicly making close to the same case: that scale alone doesn't deliver general reasoning, and that the next layer has to be causal, in direct response to recent public claims about the limits of scaling. His forthcoming textbook, Causal Artificial Intelligence, lays out that thesis at length. The overlap is worth stating plainly: this is an active, current research direction with a serious, credentialed advocate, not a fringe position — and also worth stating plainly where this work differs: Bareinboim's program addresses causal reasoning broadly, while the result here is a specific, narrower formal claim about channel non-collapsibility in state estimation architectures, proved two independent ways and checked against a reproduced empirical figure.

6. Conclusion

None of this claims that Ashby, von Neumann, or Turing specifically anticipated this architecture — that claim was checked and didn't hold up. What does hold up: a documented historical pattern of structure being traded for scale, an active research critique of that trade that hasn't gone away, a formal proof already on this site using Pearl's actual do-calculus to reintroduce structure, and a currently active researcher in Pearl's own lineage making a closely related argument in public right now. That's the honest shape of the lineage — weaker in some places than the original version of this claim, and stronger where it actually counts, because every piece of it is checkable.

References

Ashby, W. R. (1952). Design for a Brain. Chapman & Hall, London / John Wiley & Sons, New York.
Ashby, W. R. (1956). An Introduction to Cybernetics. Chapman & Hall, London.
Fodor, J. A., & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: A critical analysis. Cognition, 28(1–2), 3–71.
Newell, A., & Simon, H. A. (1976). Computer science as empirical inquiry: Symbols and search. Communications of the ACM, 19(3), 113–126.
Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems. Morgan Kaufmann.
Pearl, J. (2000). Causality: Models, Reasoning, and Inference. Cambridge University Press.
Pearl, J., Glymour, M., & Jewell, N. P. (2016). Causal Inference in Statistics: A Primer. Wiley.
Lacefield, G. S. (2026). Non-Collapsible State Separation — v3.7. dynamicsystemsarchitecture.org.
dynamicsystemsarchitecture.org (2026). Test 5 — Tuned Version: Real Result, Matches the Original Citation's Magnitude.