Notes: Five Forms of State and Line Isolation Breach / Заметки: пять форм состояния и пробой изоляции
Year: 2026
Type: article
Abstract: This chronological observation log (October 29 – November 15, 2026) documents the primary empirical episodes, failure instances, and working hypotheses behind our taxonomy of epistemic degradation in large language models. Maintained as an append-only research notebook, it traces the emergence of five intra-task failure modes—task formulation loss (Form 1), premature closure bias (Form 2), referent displacement (Form 3), unnamed verification scope (Form 4), and unverified provenance (Form 5a/5b)—alongside cross-task line isolation breach (Form 6a/6b) and out-of-taxonomy tool failures. Every entry preserves its original date and explicit epistemic boundary—[measured], [witnessed], or [hypothesis]—separating reproducible transcript counts from single-instance observations and untested mechanisms. Key records include the two-entry degradation cascade on a 114,148-character session summary, internalized English self-anchors against Form 2 in latent reasoning, header-count blindness across four model architectures during diff-audits, the short-window boundary of Form 6 across Argon II testbeds (v11, v12, v12.1), and coordinate-level verification against full web-chat JSON exports (including the multi-turn relapse trajectory of a refuted figure under Form 5). Companion deposits: compiled dossier Anatomy of Degradation (10.5281/zenodo.24266564), Deterministic Audit v2.0 (10.5281/zenodo.24278350), Attention Dispersion Law v3.0 (10.5281/zenodo.24278396), and the narrative Trilogy (10.5281/zenodo.24278460). The Russian version of this log is the second file in this record:NOTES_v1.0_ru.pdf. / Русская версия журнала — второй файл записи: NOTES_v1.0_ru.pdf.
Author: Borovikov, Vladimir V.
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234365658
