Storing differences rather than the common ground: Zhuravlev test separators for bank balance sheets and tax treaties — method note, applications and history

This note describes how I use the test (dead-end test) approach of the Yu. I. Zhuravlev school of pattern recognition in two settings: monthly balance sheets of banks and a corpus of German tax-law texts, including double taxation treaties (DTTs). A test is a set of features on which any two objects of different classes differ; a dead-end test is a test that becomes invalid when any feature is removed; the weight of a feature is the share of dead-end tests that contain it. The working principle follows from this definition: what all objects share (the background) carries no information; information lies in the deviations. Every bank has correspondent accounts, cash and a loan book; what matters is how it differs from the background. For treaties, the common part is one (the OECD Model); a country should store and search only its deviation (decided 26 October 2026, not yet built). I formulate the problem for large data: which hyperplanes divide a cloud of several hundred thousand bank-month points, and do the cuts stay stable over time. The method is classical; this note applies it and does not propose a new one. I report what is measured and published (dataset bank101: 2,157 banks, 287 dates, 39.4 million rows, 966 coordinates; a pilot of 20 dead-end tests on legal material; 380 deviations among 8,868 cost-allocation fragments) separately from preliminary results still to be re-verified. Part of a series; overview: https://doi.org/10.5281/zenodo.23949487. Files: English version (authoritative) and German version (Deutsche Fassung). Data: https://huggingface.co/datasets/sevriukov/bank101