Anatomy of Degradation: Five Forms and Line Isolation Breach / Анатомия деградации: пять форм и пробой изоляции
Year: 2026
Type: preprint
Abstract: Large language models operating across long working sessions fail not by losing fluency, but by drifting away from ground-truth artifacts while remaining smooth and confident. We classify these failures into five intra-task forms—task formulation loss (Form 1), premature narrative closure (Form 2), referent displacement (Form 3), unnamed verification scope (Form 4), and unverified provenance (Form 5)—plus a sixth cross-task mode, line isolation breach (attribute diffusion 6a and unprompted frame merging 6b), kept distinct from out-of-taxonomy tool failures. Every assertion in the dossier carries an explicit epistemic status—[measured], [witnessed], or [hypothesis]—auditing a two-entry degradation cascade across exported agent and web-chat session logs alongside prospective stress-tests from the Argon II series (v1–v1.2, v8–v8.5, and v11–v12.1). Controlled runs resolve slice-level power on taxonomy noise (K = 12, 72 calls, DiD p = 0.3555), show that filtering audit findings by verified verbatim quotes cuts recall by 61% without improving precision (64.8% vs. 63.7%, Fisher p = 0.8891) because genuine passport constants are attached to wrong subjects (Form 3), and establish that short sparse-window text testbeds do not trigger Form 6 above background (0/32 sets in v11; 1/12 = 8.3% per model in v12.1). Companion deposits: Observation Log NOTES (10.5281/zenodo.24266885), 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 dossier is the second file in this record: DEGRADATION_v1.0_ru.pdf. / Русская версия — второй файл записи: DEGRADATION_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: W7234215640
