The LeckerMath pipeline for verifiable LLM-assisted work in German tax law: overview, registry of methods, history and open problems (2024–2026)

This preprint is the overview and registry of a series. It describes a working pipeline for LLM-assisted research in German tax law that I have developed since spring 2024. External large language models supply breadth; a local base of official legal texts supplies the anchor; a human interprets and decides. Confidential case material never leaves the local machine: only anonymised questions do. The overview lists 23 methods with the date of their first dated record, the evidence for that date, the current status (measured, prototype, direction) and the detailed preprint that describes them. The methods cover structure-preserving segmentation of statutes, decisions and administrative guidance; folding of legitimate duplicates; a cascade doctrine for heterogeneous documents; evaluation by games with hidden facts and blind LLM judges; blind authoring of control questions from randomly sampled norms; answers as a space bounded by several norms, with clarifying questions; point-in-time law; a confidentiality gateway with a re-identification bench; and Zhuravlev's test separators for storing differences rather than the common ground, applied to bank balance sheets and proposed for tax treaties. Most components have close prior work, which Section 2 names; the series documents their combination for German tax law and its dated history, not a new single method. I report the main measurements with their dates, sample sizes and limits, and the open problems. The purpose of the series is to document the methods and their history in a citable, dated and versioned form. Files: English version (authoritative) and German version (Deutsche Fassung). Not legal or tax advice.