Forensic Auditing of Post-event Predictor Leakage in Spatial Wildfire Graphs: A Provenance-Controlled Evaluation of a Fiedler Embedding

Spatial graph embeddings and spectral coordinates, including Fiedler vectors (|v2|),have been proposed for landscape partitioning and spatial-connectivity analysis. Retrospectiveevaluations of spatial graph models are, however, vulnerable to implicit predictorleakage when post-event features enter edge weights or node attributes. We conducta forensic audit of spatial graph embedding representations for wildfire prediction. Previouslyreported performance in historical cohorts was confounded by post-event spatialattributes embedded in node vectors. When predictor graphs were re-exported usingstrictly pre-incident environmental attributes, the ungrounded Fiedler-vector magnitude(|v2|) did not provide consistent discrimination between burned and unburnednodes in the primary cohort eligible under the available provenance record (n = 2: LaPalma 2023, AUC = 0.307 [0.275, 0.332]; Tejeda B 2019, AUC = 0.518 [0.413, 0.594]).Simple baselines, maximum incident conductance and elevation, showed higher withincasediscrimination (AUCs 0.77–0.90). The two evaluation artefacts excluded identifiedpost-event attributes; end-to-end reconstruction from original pre-incident rasters remainspending archival verification. We establish a reproducible audit and governanceprotocol to prevent contaminated artefacts from supporting predictive claims.