The Full-Stack Architecture
of Causal Cognition

From Correlation to Causation: A Unified Framework
of Ontology · Epistemology · Methodology · Instrumentology

1Introduction: After “Correlation Does Not Imply Causation”

“Correlation does not imply causation” is a foundational doctrine of statistical training. However, this warning obscures a more important question: then what exactly is causation? How do humans identify causation within correlated information?

Pearl’s (2000) Structural Causal Models (SCM) and Rubin’s (1974) Potential Outcomes framework have rigorously developed the mathematical tools of causal inference. But these tools focus on work “after data is given,” remaining silent on the origins of data — what instruments are used for observation, who recorded what, and what characteristics the observer’s cognitive hardware possesses.

This paper presents the entire process of causal cognition as a unified layered architecture. The architecture proceeds from the physical limits of observation tools (Level -1) through to mathematical reasoning (Level 8), encompassing the entire process in a layered structure where the output of each layer serves as input to the next. The framework’s four strata — Ontology, Epistemology, Methodology, Instrumentology — are each summarized in a single sentence:

① Ontology: Objective facts and objective data exist.


② Epistemology: As cognitively capable intelligent beings, humans engage in cognition as the act of understanding objective facts and objective data. This act is a dynamic process of the subjective aligning with the objective.


③ Methodology: The primary instrument of human cognition is logical thinking, and the core of logical thinking is determining valid causal connections between objective facts and objective data — that is, causal filtering + causal ordering + causal bounding.


④ Instrumentology: Whether mathematical formulas, philosophical metaphysics, physics and chemistry formulas, law, engineering standards, medical protocols, linguistic narrative, or computational models — all are instruments for achieving closure of objective causal validity and enabling human reuse and propagation.

2The Evolution of Causal Definitions: A Chronological Survey

2.1 The Philosophical Foundation Period (1748–1911)

David Hume (1748) pointed out that humans have never directly observed causation itself; what we observe is merely the “constant conjunction” of two events. Karl Pearson (1911) pushed this position to its extreme, dismissing causation as “superstition” in science and arguing that correlation coefficients alone suffice.

2.2 The Emergence of Operational Definitions (1920s–1965)

Sewall Wright (1921) invented path analysis, becoming the first person to represent probabilistic causal relationships with directed graphs. Ronald Fisher (1935) established the experimental design principles of randomization, replication, and blocking. Austin Bradford Hill (1965) proposed nine criteria for evaluating whether an epidemiological association is causal in nature — strength, consistency, specificity, temporality, biological gradient, plausibility, coherence, experimental evidence, and analogy.

2.3 The Competition of Formal Definitions (1969–2000)

Granger (1969) proposed a predictive causal test based on temporal precedence. Rubin (1974) proposed defining causal effects using potential outcomes. Mackie (1965/1974) proposed the INUS condition definition. Pearl (1995, 2000), through Structural Causal Models (SCM) and the do-calculus, unified graphical models, potential outcomes, counterfactuals, and structural equations.

2.4 Pearl’s Causal Hierarchy and the Current State

Pearl’s Causal Hierarchy Theorem proved the mathematical impossibility of crossing between three levels — observation P(Y|X), intervention P(Y|do(X)), and counterfactuals. The 2021 Nobel Prize in Economics was awarded for causal inference methodology (natural experiments), marking the highest level of academic recognition for the methodology of distinguishing correlation from causation. However, a unified definition of causation remains absent, and causal pluralism is the current mainstream position.

3The Binary Ontology of Causation: The Primacy of Boundary Demarcation

3.1 The Relata Problem

The standard view of causal relationships is a binary relation between two relata (Stanford Encyclopedia of Philosophy). However, there is no consensus on the categories of relata (events, facts, properties, variables, processes), their individuation methods, or the adicity of the relation. Davidson’s (1967) coarse-grained event view, Kim’s (1976) ⟨object, property, time⟩ triple, and Pearl’s (2000) variable-value scheme each compete — choosing a different ontology yields entirely different causal judgments.

3.2 The Granularity Problem

The same phenomenon can be described at the molecular, cellular, individual, or population level. At which granularity a researcher defines “cause” and “effect” depends entirely on the research question and disciplinary tradition. The boundaries of local events are essentially arbitrary — igniting paper requires oxygen, temperature, humidity, Earth’s gravity, and all other conditions, yet researchers select only some of these as “causes.”

3.3 The Role of the Temporal Dimension

Hume regarded “the cause preceding the effect in time” as an essential element of causation, but Pearl’s SCM removed temporal precedence from the core definition of causation. Experiments on retrocausality and indefinite causal order in quantum physics suggest that temporal precedence may not be a necessary condition for causation. The trend in modern causal theory is to demote time from a “definition” of causation to a “diagnostic tool.”

4The Proportion of Subjective Determination in Causal Definitions

Pearl himself explicitly acknowledged: “Behind any causal conclusion there must lie some causal assumption that received no deductive or empirical support — assumptions that are untested and untestable in observational studies.” Analysis of the entire process of causal inference reveals that subjective determination intervenes at no fewer than seven levels:

Structurally, the premises of causal inference (drawing graphs, selecting variables, specifying structure) are 100% determined subjectively by humans, while the mathematical derivations built upon these premises (do-calculus, d-separation) are 100% objective. The reliability of causal inference depends entirely on the quality of those subjective premises.

5Prerequisite Infrastructure: Observation, Recording, and Cognitive Hardware

5.1 Level -1: Observation Tools — Civilization’s Hardware Ceiling

The boundary of causal questions humans can pose equals the boundary of observational capability. What cannot be seen cannot become a variable in causal inference. Every revolution in observation tools is a hardware upgrade for causal cognition: in the era of the naked eye, diseases were attributed to “miasma”; after the optical microscope (1674) made bacteria visible, the microbial theory of disease became possible; after the electron microscope (1931) made viruses visible, viral etiology could be established; genomic sequencing (1953–) pinpointed the molecular causation between specific gene mutations and specific diseases.

5.2 Level 0: Observation Behavior — Theory-Laden Perception

As Hanson (1958) systematically argued, scientific observation is by no means pure visual reception. When two scientists view the same bubble chamber photograph through fundamentally different theoretical frameworks, they do not see the same thing. The selection of what to observe (what to attend to), its conceptualization (what to call it), and its measurement (what instrument to use) are all influenced by theoretical presuppositions.

5.3 Level 0.1: Data Recording — Externalized Memory

The working memory of the human brain is extremely limited, and long-term memory is highly unreliable. The core demands of causal inference — precise comparison of multiple observations, long-duration data alignment, and large-sample statistical pattern recognition — exceed the brain’s unaided capacity. Recording transforms “one-time personal experience” into “repeatedly reviewable public data.”

The case of John Graunt (1662) is highly illustrative. London’s death registers had been published weekly since 1603, yet for 60 years no one extracted causal information from them. It was not until the haberdasher Graunt analyzed 70 years of data that urban–rural mortality differentials were discovered, epidemics were distinguished from endemic diseases, and the first human life table was compiled. Data does not automatically generate insight, but without data, insight cannot possibly emerge.

5.4 Level -2: Cognitive Hardware — The Biological Ceiling

When the same apple falls from a tree, Newton’s brain produced a “sense of surprise,” while his neighbor felt nothing. This is not an educational difference (installable software), but rather individual variation in cognitive hardware: sensory thresholds, working memory capacity, pattern recognition sensitivity, surprise thresholds, and abstraction capability. Behavioral genetics research indicates that the heritability of cognitive ability (the g factor) is approximately 0.5 to 0.8 in adulthood. The upper limit of causal cognitive capability is ultimately determined by biologically endowed hardware parameters.

6Logical Thinking = Causal Filtering + Ordering + Bounding

6.1 Redefining Peirce’s Three Logics as Causal Operations

Charles Sanders Peirce (1903) reconceived deduction, induction, and abduction not as three parallel types of reasoning but as three consecutive stages of scientific inquiry. This paper argues that all three logics are different modalities of causal operations:

Under this redefinition, all logical fallacies reduce to the failure of one step among the three operations: the post hoc fallacy is a filtering failure (treating non-causal as causal); reversing cause and effect is an ordering failure (drawing the arrow in the wrong direction); overgeneralization is a bounding failure (drawing boundaries too broadly).

6.2 The Asymmetry of Prerequisites

All three logics require “observed facts” as input. Peirce himself explicitly identified the starting point of abduction as “the surprising fact C is observed.” But this single sentence presupposes that Level -1 (observation tools), Level 0 (observation behavior), and Level 0.1 (data recording) have all been completed. Logical thinking can operate only atop this prerequisite infrastructure.

7The Causal Cognition Process of “Subjective Aligning with the Objective”

Distinguishing data that possesses causal relationships from data that does not, within correlated information, is a critical process of human subjective alignment with the objective. This proposition simultaneously contains three layers of meaning:

First, it acknowledges the existence of the subjective. Humans are not passive receptacles for objective causal relationships but use subjective cognitive hardware to approximate objective structures. Second, it acknowledges the existence of the objective. Although causal judgment is subjective, the thing it attempts to align with — the causal structure of the world — is objective. Fire truly burns paper; it is not merely our subjective belief that fire burns paper. Third, it acknowledges that alignment is a process. It is not completed in a single instance but through continuous cycles of observation → recording → filtering → testing → revision, spiraling toward approximation.

Taken together, the three layers precisely describe the essence of science — neither pure objectivism (“data speaks for itself”) nor pure subjectivism (“everything is socially constructed”), but a dynamic process in which the subjective continuously calibrates toward the objective.

8Instrumentology: Eight Instruments of Causal Closure

If causal judgments exist only in individual minds, they dissipate with the individual’s death. For causal knowledge to become an asset of civilization, instruments are needed that achieve closure of causal judgments — that is, encapsulating them in forms that others can independently verify and reuse.

An inverse relationship exists between precision and propagation. F=ma is extremely precise but comprehensible only to those with education; “you reap what you sow” is imprecise but universally understood. The knowledge propagation system of human civilization is the process of balancing between these two poles.

9The Convergence of Eastern and Western Theories of Causation: Civilizational Difference or Individual Difference

The co-author of this paper — the human researcher — is a Korean Buddhist practitioner. His core judgment is as follows: the differences between Eastern and Western philosophy and science in their understanding of the nature of causation fall fundamentally within the range of inter-individual cognitive differences and do not constitute a systematic civilizational divergence.

The evidence supporting this judgment: Hume’s (1748) “constant conjunction” and Nāgārjuna’s (~150 CE) “all phenomena arise from conditions; all phenomena perish from conditions” are expressions of the same insight in different languages. Peirce’s (1903) “starting from the surprising fact” and the Buddha’s “this being, that becomes” are products of the same cognitive structure. Pearl’s (2000) “causal assumptions cannot be verified by data” and the Yogācāra school’s “all is mind-only” point to the same limitation.

They arrive at similar conclusions not because they belong to different civilizations, but because, as equally perceptive individuals of the same species, they collided with the same wall. The ceiling of causal cognition is not drawn between civilizations but above the biological limits of human cognitive capacity.

10Unified Architecture: The Full-Stack Diagram

11Blind Spot Analysis

A self-critical examination of this framework identifies the following blind spots: (1) Social construction — the power structure determining whose observations are recognized as legitimate material by the academic community has not been incorporated as an independent level. (2) The invisible filter of language — a variable can be named only because the language already contains a corresponding word; this linguistic prerequisite has not been independently discussed. (3) Conflation of observation and intervention — passive observation and experimental intervention have not been rigorously distinguished. (4) The origins of counterfactual imagination have not been explored. (5) The propagation chain from “being discovered” to “being accepted” to “being acted upon” for causal knowledge has not been analyzed. (6) The violence of discretization upon continuity — the discretization of “cause” and “effect” itself may be a violent simplification of continuous reality. (7) Ethical ceiling — a large number of causal relationships involving humans are in principle impossible to verify through interventional experiments. (8) The self-referential problem of the framework itself — as a causal claim, this architecture cannot fully verify its own premises.

12Conclusion

This paper integrates the entire process of causal cognition into a layered architecture. The core conclusions are summarized in four sentences:

Ontology: Objective facts and objective data exist.


Epistemology: As cognitively capable intelligent beings, humans engage in cognition as the act of understanding objective facts and objective data. This act is a dynamic process of the subjective aligning with the objective.


Methodology: The primary instrument of human cognition is logical thinking, and the core of logical thinking is determining valid causal connections between objective facts and objective data.


Instrumentology: Whether mathematical formulas, philosophical metaphysics, physics and chemistry formulas, law, engineering standards, medical protocols, linguistic narrative, or computational models — all are instruments for achieving closure of objective causal validity and enabling human reuse and propagation.

The entire accumulated knowledge of human civilization is the process of filtering out causal structures from ever-greater amounts of correlation, confirming their direction and boundaries, encapsulating them in reusable forms through closure instruments, and passing them on to the next generation. This is the definition of science, the definition of education, and the definition of civilization itself.

Ref.References