Code and experiment archives for "Enforcing Graph-Relative Structural Compliance in LLM-Based Industrial Anomaly Explanation via Causal Graph Grounding"

Le résumé fourni par la source

Code and experiment archives for the manuscript "Enforcing Graph-Relative Structural Compliance in LLM-Based Industrial Anomaly Explanation via Causal Graph Grounding". The package contains the full research pipeline: a synthetic SCM generator; the PCMCI causal-discovery pipeline with exact-reproduction scripts for the SWaT and WADI discovery runs (51->28 retained dimensions, 230 cross-variable edges); an LLM backend implementing edge-set grounding (enum hard constraint via JSON-schema structured output with client-side validation, plus soft-prompt variants); run scripts for all reported experiments, including 30-seed degree-matched random-graph controls, the honest-boundary audit, evidence-table injection, and four-vendor cross-replication (Alibaba, DeepSeek, Zhipu, Moonshot); per-run JSON and log archives; and the raw per-rater CSV files of the human study (14 raters). The synthetic benchmark is self-contained; SWaT and WADI require a licence from iTrust, SUTD and are not redistributed. All seeds are fixed; see README.md for entry points and results/RUNS_2026_METADATA.md for model snapshot identifiers. License: MIT.Version 2 adds per-seed experiment archives and scripts E32-E41 (evidence-table CI re-run, rewire-overlap audit, agreement coefficients, structure-fault arms, allowed-set coverage with wrong-candidate sensitivity, honest-boundary reconstructions, cap-policy and prompt-metric comparisons, size-matched correlation control, and seed extensions). Version 3 accompanies the revised, retitled Expert Systems submission: it adds the WADI variable re-selection with unit-invariant criteria and the first WADI event-level evaluation (E42), a graph-drift probe (E43), structured numeric-claim fields (E44), an abstention 15-window extension (E45), the WADI unit-invariant compliance arm for the main table (E46), a root-incident-edge true-gap probe (E47), a unified WADI A1/A2 pipeline (E48), the extended WaDi.A2 release evaluation (E49), and the revised manuscript sources.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.