Tensor World-State, Verifiable Orchestration, Topological Terrain Descriptors and Optimal-Transport Contextual Similarity

Abstract

While LLM-based tool orchestration (sometimes referred to as agentic tools) is an emerging trend in remote sensing analysis, current versions lack outputs that can be verified, low tool-orchestration error rates and any principled way to deal with hyperspectral data. The following study presents the multi-agent system TerraXIA that is used for the analysis of hyperspectral images of terrains. The language model does not analyze, it only plans and interprets, while all analysis is done by deterministic mathematical operators and only what an operator calculated can be stated by an agent. A total of four contributions has been made. First, a world-state (N1) describes the spectral state of one or many tensors with Tucker decomposition and limited unmixing is used to create a compressed, interpretable substrate, reconstruction error 2.5%, compared to ground-truth, compression by a factor 19.5, endmember spectral angle 0.149 rad and abundance RMSE 0.156. Second, a verifiable orchestration architecture (N2) where every piece of output from every operator is a content-hashed and spatially indexed artefact and every report claim has to cite an artefact. When the operator registry was frozen into the model, the LLM planner was successful at generating statically valid plans on the 20 test intents with no tool-call error and the grounding checker disambiguated another 42% of calls reported through the first pass of the model but if the operator registry was not in the model, the grounding checker would have rejected all the generated plans and the 53% of the plans would have been statically invalid with error in calls. Third, a persistent-homology terrain descriptor (N3) having all topological features remapped back to georeferenced pixel polygons and which changes only around a lower rate of 25,000 times compared to a GLCM texture baseline during rotation, flip and monotone illumination change. Fourth, an optimal-transport contextual-similarity operator (N4) that is competitive with, but not outperforming more basic spectral baselines in the current retrieval and anomaly tasks and for which the metric properties are expressed numerically with machine precision and clearly explained. All reported numbers can be reproduced from the artifact store that is hash-addressed.