Systems, Vol. 13, Pages 784: LLM-Powered, Expert-Refined Causal Loop Diagramming via Pipeline Algebra

Systems doi: 10.3390/systems13090784

Authors:
Kirk Reinholtz
Kamran Eftekhari Shahroudi
Svetlana Lawrence

Building a causal-loop diagram (CLD) is central to system-dynamics modeling but demands domain insight, the mastery of CLD notation, and the ability to juggle AI, mathematical, and execution tools. Pipeline Algebra (PA) reduces that burden by treating each step—LLM prompting, symbolic or numeric computation, algorithmic transforms, and cloud execution—as a typed, idempotent operator in one algebraic expression. Operators are intrinsically idempotent (implemented through memoization), so every intermediate result is re-used verbatim, yielding bit-level reproducibility even when individual components are stochastic. Unlike DAG (directed acyclic graph) frameworks such as Airflow or Snakemake, which force analysts to wire heterogeneous APIs together with glue code, PA’s compact notation lets them think in the problem space, rather than in workflow plumbing—echoing Iverson’s dictum that “notation is a tool of thought.” We demonstrated PA on a peer-reviewed study of novel-energy commercialization. Starting only from the article’s abstract, an AI-extracted problem statement, and an AI-assisted web search, PA produced an initial CLD. A senior system-dynamics practitioner identified two shortcomings: missing best-practice patterns and lingering dependence on the problem statement. A one-hour rewrite that embedded best-practice rules, used iterative prompting, and removed the problem statement yielded a diagram that conformed to accepted conventions and better captured the system. The results suggest that earlier gaps were implementation artifacts, not flaws in PA’s design; quantitative validation will be the subject of future work.

Source link

Kirk Reinholtz www.mdpi.com