Work

Tawjih: Directing a Language Model with Three-Sided Files and Cross-Tree Production Paths

Year: 2026
Type: article
Abstract: Working paper describing version 0.6.1 of Tawjih, a file-based method for directing a language model from organized personal knowledge, with a reference implementation in a single Python file that uses only the standard library. Knowledge is kept in trees of small files. Each file has three sides: the human concept, machine language (instructions for the model and Python code), and a reading side holding passages filtered from deposited books. At build time the reading side is turned into the concept and the concept into machine language; at run time only machine language is sent. A path connects files from any tree with the words TO, AND, IF, AMONG and THROUGH. Version 0.6.1 adds data contracts between steps, detection of machine language that is older than its sources, a cache of model answers so a failed run can resume, drawing and generation of paths, a time limit for code steps, and the reading of books from their source formats (Markdown, LaTeX, HTML and PreTeXt XML) so that mathematics survives. The paper reports a trial on two chapters of an open algebra textbook (Judson, Abstract Algebra: Theory and Applications). The author's assistant played the language model through a manual back end, so the trial tests the program and the workflow, not the quality of a model. It exposed six defects, all fixed in this version, and showed that text extracted from the PDF damaged the mathematics while the book's PreTeXt source preserved it. The implementation passes 97 automated tests on Linux and Windows. No evaluation with a small local model has been made yet. Source code: https://github.com/AlmodinyMk/LLM-tawjih-Model-building
Language: English
Open Access status: green
OpenAlex ID: W7236217199
FWCI:
Cites:
Cited by:
Related to: 10