Inhalt
Aktueller Ordner:
/ARS_MetaMeth_Eng.tex
% English Version
\documentclass[12pt,a4paper]{article}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage{xcolor}
\usepackage{hyperref}
\usepackage{geometry}
\geometry{a4paper, left=3cm, right=3cm, top=3cm, bottom=3cm}
\usepackage{setspace}
\onehalfspacing
\usepackage{parskip}
\usepackage[english]{babel}
\usepackage{csquotes}
\usepackage{microtype}
\usepackage{booktabs}
\usepackage{longtable}
\usepackage{array}
\usepackage{listings}
\usepackage{caption}
\usepackage{subcaption}
\usepackage{float}
\usepackage{url}
\usepackage{natbib}
\usepackage{titling}
\lstset{
language=Python,
basicstyle=\ttfamily\small,
keywordstyle=\color{blue},
commentstyle=\color{green!40!black},
stringstyle=\color{red},
showstringspaces=false,
numbers=left,
numberstyle=\tiny,
numbersep=5pt,
breaklines=true,
frame=single,
backgroundcolor=\color{gray!5},
tabsize=2,
captionpos=b
}
\title{\Huge\textbf{ARS as a Meta-Methodology} \\[2mm]
\LARGE The Conditions for Explanatory Models \\[2mm]
\LARGE in the Age of Generative AI}
\author{
\large
\begin{tabular}{c}
Paul Koop
\end{tabular}
}
\date{\large 1994--2026}
\begin{document}
\maketitle
\begin{abstract}
This paper examines the Algorithmic Recursive Sequence Analysis (ARS) not merely
as a method but as a \textit{meta-methodology}—a framework that specifies the
conditions under which a model counts as \textit{explanatory} rather than merely
\textit{descriptive} or \textit{simulative}. Drawing on the ARS's core principles—the
primacy of interpretation, the separation of structure and statistics, controlled
falsification, and XAI validation—I argue that these principles constitute necessary
criteria for explanatory models in any discipline that deals with sequential social
processes. The paper systematically relates ARS to five contemporary research
programs: (1) Formal Verification and Model Checking, (2) Interpretable Machine
Learning and Rule Extraction, (3) Grounded Theory, (4) Causal Inference, and
(5) Process Mining. For each, I demonstrate what ARS can learn from these
approaches and, crucially, what these approaches can learn from ARS. The paper
concludes that ARS provides a transdisciplinary benchmark for distinguishing
genuine explanation from statistical or structural description.
\end{abstract}
\newpage
\tableofcontents
\newpage
\section{Introduction: From Method to Meta-Methodology}
The Algorithmic Recursive Sequence Analysis (ARS), in its various versions
(2.0–4.0), has been presented primarily as a method for the formal analysis
of sequential interactions. It transforms interpretively obtained categories
into formal models: probabilistic context-free grammars (PCFG), Petri nets,
Bayesian networks, and deterministic finite automata (DFA). This paper takes
a different perspective. It argues that ARS is not just a method but a
\textit{meta-methodology}—a framework that specifies the \textit{conditions}
under which a model can be considered explanatory.
This shift in perspective is motivated by a recurring confusion in contemporary
AI and data science: the conflation of \textbf{simulation} with \textbf{explanation}.
A large language model (LLM) trained on sales conversations can simulate
plausible dialogues with high fidelity. Yet, as the ARS notebooks have
demonstrated, it cannot explain \textit{why} a particular sequence of speech
acts is well-formed or what rules constitute its generation. The LLM provides
a statistical shadow of the process; the ARS provides its structural skeleton.
The meta-methodological claim of this paper is that the ARS principles
constitute \textbf{necessary conditions} for explanatory models in any
discipline dealing with sequential social processes. These conditions are:
\begin{enumerate}
\item \textbf{Interpretive Grounding}: Symbols must be tied to documented
qualitative interpretations, not merely to statistical correlations.
\item \textbf{Structural Decidability}: The well-formedness of sequences
must be formally decidable (e.g., by a DFA), independent of empirical
frequencies.
\item \textbf{Generative Transparency}: The model must be able to generate
sequences in a way that every step is traceable to explicit rules.
\item \textbf{Falsifiability}: Interpretations must be subject to
controlled falsification; counterexamples must be able to refute rules.
\item \textbf{XAI Validation}: The model must satisfy the NIST XAI
criteria—meaningfulness, accuracy, and knowledge limits.
\end{enumerate}
To substantiate this claim, I systematically relate ARS to five contemporary
research programs that share overlapping concerns: formal verification,
interpretable machine learning, grounded theory, causal inference, and process
mining. For each, I ask two reciprocal questions:
\begin{enumerate}
\item What can ARS learn from this approach? (Technical or conceptual
enhancements)
\item What can this approach learn from ARS? (Methodological safeguards,
criteria for explanation)
\end{enumerate}
\section{ARS as a Meta-Methodology: The Core Principles}
Before examining the five approaches, it is necessary to state the ARS
principles that serve as the meta-methodological benchmark.
\subsection{Principle 1: Interpretive Grounding}
In ARS, every terminal symbol (KBG, VBBd, KAA, etc.) is the product of a
documented qualitative interpretation. The coding process is not a black box;
it is recorded, justified, and subject to intersubjective validation. This
principle excludes purely data-driven category formation (e.g., clustering
embeddings) from counting as explanation.
\subsection{Principle 2: Separation of Structure and Statistics}
As developed in \texttt{ARS\_XAI\_Aut2\_Ger.tex}, ARS maintains a strict
separation between structural rules (which are deterministic and context-free)
and statistical regularities (which are empirical and contingent). A structural
rule holds or does not hold—independent of how often it is violated. This
separation is absent in purely statistical models.
\subsection{Principle 3: Generative Transparency}
The induced grammar must be able to generate sequences in a traceable manner.
The transducer in Lisp, the parser in Pascal, and the inductor in Scheme each
provide a different window into this transparency. A model that cannot generate
exemplars from its own rules is not explanatory.
\subsection{Principle 4: Controlled Falsification}
Interpretations are not asserted once and for all. They are produced as
readings and then falsified by subsequent sequence positions (following
Oevermann's sequential analysis). This creates a spiral of interpretation and
refutation that mirrors Popperian falsificationism adapted to qualitative
material.
\subsection{Principle 5: XAI Validation}
The three NIST XAI criteria—meaningfulness, accuracy, knowledge limits—are
not optional add-ons but constitutive elements of explanatory models.
Meaningfulness requires semantic interpretability; accuracy requires
correspondence with the material; knowledge limits require explicit
documentation of the model's boundaries.
\section{Five Approaches in Dialogue with ARS}
\subsection{Formal Verification and Model Checking}
\subsubsection{What Formal Verification Is}
Formal verification, particularly model checking, is a method from theoretical
computer science that systematically checks whether a formal model (e.g., a
finite automaton, a Petri net, or a Bayesian network) satisfies specified
properties. These properties include \textit{safety} ("something bad never
happens") and \textit{liveness} ("something good eventually happens").
Model checking is exhaustive: it explores the entire state space of the model.
Unlike statistical testing, which provides probabilistic guarantees, model
checking provides \textit{deterministic} guarantees about the model's behavior.
\subsubsection{What ARS Can Learn from Formal Verification}
\begin{itemize}
\item \textbf{Property Specification Languages}: ARS could adopt temporal
logics (LTL, CTL) to specify what properties the induced grammar should
satisfy. For example: $\square (KBG \rightarrow \lozenge VBG)$ ("Always,
if a customer greets, eventually the seller greets back").
\item \textbf{Counterexample Generation}: When a property fails, model
checkers produce a counterexample trace. This could serve as a powerful
falsification tool for ARS interpretations.
\item \textbf{State Space Explosion Awareness}: ARS modelers should be
aware that hierarchical grammars can lead to large state spaces. Formal
verification offers abstraction techniques to manage this.
\end{itemize}
\subsubsection{What Formal Verification Can Learn from ARS}
\begin{itemize}
\item \textbf{Interpretive Grounding of States}: In standard model checking,
states are abstract symbols. ARS insists that each state must be
interpretively grounded. This could lead to a new subfield: \textit{interpretive
model checking}, where properties are checked \textit{and} the meaning of
states is documented.
\item \textbf{The Separation of Structure and Statistics}: Model checking
typically assumes a deterministic model. ARS shows how to separate
structural rules (which can be verified) from statistical variations
(which cannot). This could enrich probabilistic model checking.
\item \textbf{XAI Criteria for Verification}: Model checking results are
often opaque to non-experts. ARS's XAI criteria could guide the development
of more understandable verification outputs.
\end{itemize}
\subsection{Interpretable Machine Learning and Rule Extraction}
\subsubsection{What IML and Rule Extraction Are}
Interpretable Machine Learning (IML) aims to make black-box models (neural
networks, gradient boosting, random forests) understandable to humans.
\textit{Rule extraction} is a specific IML technique that attempts to describe
the learned function approximately through a set of if-then rules (e.g., using
RIPPER, CART, or analyzing activation patterns).
Unlike ARS, which induces rules directly from data, rule extraction typically
works \textit{post-hoc}: the model is already trained, and rules are extracted
as an explanation.
\subsubsection{What ARS Can Learn from IML and Rule Extraction}
\begin{itemize}
\item \textbf{Scalable Rule Induction}: ARS currently induces rules from
small corpora (n = 8). IML offers techniques for extracting rules from
large datasets, albeit with less interpretive control.
\item \textbf{Quantitative Rule Evaluation}: IML provides metrics for
evaluating extracted rules (coverage, fidelity, stability). ARS could
adopt these to assess how well a grammar generalizes.
\item \textbf{Hybrid Rule Sets}: Some IML methods combine global and local
rules. ARS could explore hybrid grammars that have both a core structural
grammar and local statistical variations.
\end{itemize}
\subsubsection{What IML and Rule Extraction Can Learn from ARS}
\begin{itemize}
\item \textbf{Explication vs. Approximation}: IML's rule extraction is
almost always approximate. ARS insists on \textit{exact} rules for the
given corpus. This raises a fundamental question: Is approximation ever
acceptable for explanation? ARS suggests a clear answer: approximation
is acceptable only if the approximation error is documented and the
structural core is exact.
\item \textbf{Interpretive Validation of Rules}: IML extracts rules that
statistically fit the data. ARS adds a layer of \textit{interpretive
validation}: rules must also make sense to human interpreters. This could
prevent the extraction of statistically correct but semantically meaningless
rules.
\item \textbf{Falsifiability as a Criterion}: IML rarely discusses how
extracted rules could be falsified. ARS makes falsifiability a central
criterion. IML could adopt this: a rule set is not explanatory if no
conceivable counterexample could refute it.
\end{itemize}
\subsection{Grounded Theory}
\subsubsection{What Grounded Theory Is}
Grounded Theory (GT), developed by Glaser and Strauss, is a classic methodology
in qualitative social research for developing theories from data. It involves
procedures such as open coding, axial coding, and selective coding. The goal
is to generate middle-range theories that are "grounded" in empirical material.
In recent years, there have been attempts to formalize parts of GT or to
support it computationally (e.g., through natural language processing or
topic modeling). However, GT remains largely informal in its final output.
\subsubsection{What ARS Can Learn from Grounded Theory}
\begin{itemize}
\item \textbf{Systematic Coding Procedures}: GT offers a rich vocabulary
and set of procedures for coding that could enrich ARS's interpretive
phase. Concepts like "axial coding" (relating categories to subcategories)
are similar to ARS's hierarchical compression but more fine-grained.
\item \textbf{Theoretical Sampling}: GT's principle of theoretical
sampling—selecting new cases based on emerging theoretical insights—could
guide ARS's case selection in larger studies.
\item \textbf{Constant Comparison}: GT's method of constant comparison
(comparing each new case with already developed categories) is already
implicit in ARS's systematic case comparison (Phase 4) but could be
made more explicit.
\end{itemize}
\subsubsection{What Grounded Theory Can Learn from ARS}
\begin{itemize}
\item \textbf{From Theory to Generative Model}: GT typically stops at the
level of narrative theory or category systems. ARS goes further: it
transforms the theory into a \textit{generative grammar} that can produce
new sequences. GT could adopt this to make its theories testable and
executable.
\item \textbf{Formal Falsifiability}: GT's validation procedures are
primarily qualitative (e.g., member checking, peer debriefing). ARS adds
formal falsifiability: the grammar can be wrong in a way that can be
demonstrated mechanically (e.g., by a parser rejecting a sequence).
\item \textbf{XAI Criteria for Grounded Theory}: ARS's XAI criteria
(meaningfulness, accuracy, knowledge limits) provide a checklist for
evaluating grounded theories. A GT theory that cannot specify its
knowledge limits is incomplete.
\end{itemize}
\subsection{Causal Inference and Causal Graphical Models}
\subsubsection{What Causal Inference Is}
Causal inference, particularly with causal graphical models (e.g., DAGs,
DoWhy, CausalNex), goes beyond mere correlation and attempts to identify
and quantify causal relationships between variables. It uses techniques
such as the do-calculus, instrumental variables, and counterfactual reasoning.
A key insight of causal inference is that correlation is not causation.
Directed acyclic graphs (DAGs) are used to represent assumptions about
causal structure.
\subsubsection{What ARS Can Learn from Causal Inference}
\begin{itemize}
\item \textbf{Causal Interpretation of Grammars}: ARS grammars describe
sequential dependencies. Causal inference could help distinguish whether
these dependencies are merely sequential or genuinely causal. For example:
Does the question "Anything else?" \textit{cause} an additional purchase,
or is it merely correlated with it?
\item \textbf{Counterfactual Reasoning}: Causal inference excels at
answering counterfactual questions ("What would have happened if the
seller had not asked?"). ARS could adopt counterfactual simulation
(already present in Phase 3 of CGTI) as a standard validation tool.
\item \textbf{Instrumental Variables for Sequential Data}: ARS deals with
sequential data where confounding is common. Instrumental variable
techniques could help identify causal effects even in observational
sequential data.
\end{itemize}
\subsubsection{What Causal Inference Can Learn from ARS}
\begin{itemize}
\item \textbf{Interpretive Grounding of Causal Graphs}: In standard causal
inference, the DAG is often assumed or learned from data without interpretive
documentation. ARS insists that every node and edge must be interpretively
grounded. This could lead to \textit{interpretive causal inference} as a
new subfield.
\item \textbf{Sequential Grammars as Causal Structures}: ARS grammars are
a form of causal structure over sequences. Causal inference typically
deals with static or time-series data, not with grammatical sequences.
ARS could inspire a new class of \textit{grammatical causal models}.
\item \textbf{XAI for Causal Models}: Causal models are often presented
as DAGs with probabilities, which are not self-explanatory. ARS's XAI
criteria could guide the documentation of causal models, making them
more accessible to domain experts.
\end{itemize}
\subsection{Process Mining}
\subsubsection{What Process Mining Is}
Process mining is a research field at the intersection of data mining, machine
learning, and process modeling. It aims to discover, conform, and enhance
process models (often in the form of Petri nets, BPMN diagrams, or directly
follows graphs) from event logs—sequential recordings of process steps, e.g.,
in workflow management systems or ERP systems.
Process mining typically works with large, anonymized, and weakly annotated
logs. It discovers the "average process" or the "most common paths." It does
not aim for a complete reconstruction of a single case's constitutive rules.
\subsubsection{What ARS Can Learn from Process Mining}
\begin{itemize}
\item \textbf{Scalable Discovery Algorithms}: Process mining offers
sophisticated algorithms (e.g., Alpha miner, Heuristics miner, Inductive
miner) for discovering Petri nets from large logs. ARS could adopt or
adapt these for larger corpora while preserving interpretive control.
\item \textbf{Conformance Checking}: Process mining includes techniques
for checking whether an event log conforms to a given model. This could
be used to validate ARS grammars against new data.
\item \textbf{Performance Analysis}: Process mining adds performance
dimensions (time, cost, frequency). ARS could be extended to incorporate
temporal and resource dimensions more systematically.
\end{itemize}
\subsubsection{What Process Mining Can Learn from ARS}
\begin{itemize}
\item \textbf{Interpretive Discovery for Small Logs}: Process mining
typically requires large logs to produce reliable models. ARS shows how
to discover models from a single case (n = 1) through interpretive
depth. This could be valuable for process mining in domains where data
is scarce (e.g., medical procedures, legal cases).
\item \textbf{Documentation of Discovery Decisions}: Process mining
algorithms make many decisions (e.g., which paths to include, how to
handle noise). These decisions are rarely documented in an interpretively
accessible way. ARS's reflexive documentation could serve as a model.
\item \textbf{Separation of Structure and Statistics}: Process mining
often produces models that mix structural rules with statistical noise.
ARS's strict separation could improve the quality of discovered models
by clearly distinguishing what is structurally necessary from what is
merely empirically frequent.
\item \textbf{XAI for Process Mining}: The models produced by process
mining (e.g., spaghetti-like Petri nets) are often hard to understand.
ARS's XAI criteria could guide the development of more explainable
process mining outputs.
\end{itemize}
\section{Toward a Transdisciplinary Benchmark for Explanation}
The five comparisons above reveal a common pattern. Each contemporary
approach has technical strengths that could enhance ARS: scalability
(IML, Process Mining), formal rigor (Verification, Causal Inference),
and systematic coding procedures (Grounded Theory). Conversely, each
approach lacks some of the meta-methodological safeguards that ARS
provides: interpretive grounding, structural decidability, generative
transparency, controlled falsification, and XAI validation.
This symmetry suggests that ARS is not merely one method among others
but a \textbf{transdisciplinary benchmark} for what counts as an
explanatory model.
\begin{table}[H]
\centering
\caption{ARS as a Transdisciplinary Benchmark}
\label{tab:benchmark}
\begin{tabular}{@{} p{3cm} p{4cm} p{6cm} @{}}
\toprule
\textbf{Criterion} & \textbf{Question} & \textbf{Absence indicates...} \\
\midrule
Interpretive Grounding & Are the symbols meaningfully documented? & Statistical correlation without understanding \\
Structural Decidability & Is well-formedness formally decidable? & Probabilistic guessing rather than rule-following \\
Generative Transparency & Can the model generate exemplars traceably? & Simulation without explanation \\
Controlled Falsification & Can counterexamples refute rules? & Unfalsifiable post-hoc storytelling \\
XAI Validation & Are meaningfulness, accuracy, and knowledge limits documented? & Technical sophistication without epistemic accountability \\
\bottomrule
\end{tabular}
\end{table}
\subsection{The Explanation vs. Simulation Distinction Revisited}
The core distinction that emerges from this analysis is between
\textbf{explanation} and \textbf{simulation}. A model \textit{simulates}
if it reproduces the statistical properties of the data. A model
\textit{explains} if it specifies the constitutive rules that generate
the phenomenon.
\begin{itemize}
\item \textbf{Simulation} is sufficient for prediction. An LLM that
accurately predicts the next token is a good simulator.
\item \textbf{Explanation} is necessary for understanding, intervention,
and normative evaluation. An ARS grammar that specifies the rules of
a sales conversation is an explanation.
\end{itemize}
The five criteria above are the conditions under which a model qualifies
as an explanation rather than merely a simulation.
\subsection{The Role of the Human Interpreter}
A recurring theme across all five comparisons is the role of the human
interpreter. In ARS, the human is constitutive: interpretation is a human
act that cannot be fully automated. In the other approaches, the human
is often external—designing algorithms, providing training data, evaluating
outputs.
This is not a weakness of ARS but a strength. ARS makes explicit what is
often implicit: that explanation is a human practice, not a property of a
model considered in isolation. A model is explanatory \textit{for someone}
who can understand it, use it, and be held accountable for it.
\section{Conclusion and Outlook}
This paper has argued that the Algorithmic Recursive Sequence Analysis
(ARS) is not merely a method but a meta-methodology—a framework that
specifies the conditions under which a model counts as explanatory.
Five core principles were identified: interpretive grounding, structural
decidability, generative transparency, controlled falsification, and XAI
validation.
The paper then systematically related ARS to five contemporary research
programs: formal verification, interpretable machine learning, grounded
theory, causal inference, and process mining. For each, I demonstrated
what ARS can learn from these approaches (technical enhancements) and,
crucially, what these approaches can learn from ARS (methodological
safeguards, criteria for explanation).
The meta-methodological claim is that the ARS principles constitute
necessary conditions for explanatory models in any discipline dealing
with sequential social processes. They provide a transdisciplinary
benchmark for distinguishing genuine explanation from statistical or
structural description.
For future research, three directions are particularly promising:
\begin{enumerate}
\item \textbf{Implementation of hybrid systems}: Integrate ARS
grammars with model checkers, rule extractors, or process mining
algorithms while preserving the meta-methodological safeguards.
\item \textbf{Empirical testing of the benchmark}: Apply the five
criteria to existing models in different disciplines and assess
whether they predict perceived explanatory quality.
\item \textbf{Extension to non-sequential domains}: While ARS was
developed for sequential data, the meta-methodological principles
may generalize to other types of models (e.g., classification,
clustering, regression).
\end{enumerate}
In conclusion: The question is not whether a model fits the data.
Statistical fit is necessary but not sufficient. The question is whether
the model meets the meta-methodological criteria that make explanation
possible. ARS provides a concrete, operationalized answer.
\newpage
\begin{thebibliography}{99}
\bibitem[Baier \& Katoen(2008)]{baier2008principles}
Baier, C., \& Katoen, J.-P. (2008). \textit{Principles of Model Checking}.
MIT Press.
\bibitem[Glaser \& Strauss(1967)]{glaser1967discovery}
Glaser, B. G., \& Strauss, A. L. (1967). \textit{The Discovery of Grounded
Theory: Strategies for Qualitative Research}. Aldine.
\bibitem[Koop(1992)]{koop1992parser}
Koop, P. (1992). \textit{Demo-Parser Chart-Parser Version 1.0}. Pascal source code.
\bibitem[Koop(1994)]{koop1994scheme}
Koop, P. (1994). \textit{Grammatikinduktion empirisch gesicherter
Verkaufsgespräche}. Scheme source code.
\bibitem[Koop(1994)]{koop1994lisp}
Koop, P. (1994). \textit{Sequenzanalyse empirisch gesicherter
Verkaufsgespräche}. Lisp source code.
\bibitem[Koop(2023)]{koop2023notebook}
Koop, P. (2023). \textit{Qualitative Sozialforschung und Große Sprachmodelle}.
Jupyter Notebook.
\bibitem[Koop(2024)]{koop2024ars}
Koop, P. (2024/2026). \textit{Between Interpretation and Computation:
Algorithmic Recursive Sequence Analysis as a Bridge between Qualitative
Hermeneutics and Formal Modeling}. the-last-freedom.org.
\bibitem[Koop(2026)]{koop2026causal}
Koop, P. (2026). \textit{Social Structures and Processes: Causal Inference
with Probabilistic Context-Free Grammars and Bayesian Networks}.
the-last-freedom.org.
\bibitem[Molnar(2022)]{molnar2022interpretable}
Molnar, C. (2022). \textit{Interpretable Machine Learning: A Guide for
Making Black Box Models Explainable}. leanpub.com.
\bibitem[Oevermann et al.(1979)]{oevermann1979methodology}
Oevermann, U., Allert, T., Konau, E., \& Krambeck, J. (1979). The methodology
of objective hermeneutics. In H.-G. Soeffner (Ed.), \textit{Interpretative
Procedures in the Social and Text Sciences} (pp. 352-434). Metzler.
\bibitem[Ortigossa et al.(2024)]{ortigossa2024xai}
Ortigossa, E. S., Gonçalves, T., \& Nonato, L. G. (2024). Explainable
Artificial Intelligence (XAI)—From Theory to Methods and Applications.
\textit{IEEE Access}, 12, 80799-80846.
\bibitem[Pearl(2009)]{pearl2009causality}
Pearl, J. (2009). \textit{Causality: Models, Reasoning, and Inference}
(2nd ed.). Cambridge University Press.
\bibitem[van der Aalst(2016)]{vanderaalst2016process}
van der Aalst, W. M. P. (2016). \textit{Process Mining: Data Science in
Action} (2nd ed.). Springer.
\end{thebibliography}
\end{document}