Written by AI, Managed by AI: Semantic Space Control and Index Sickness Elimination Across 391 Consecutive Sessions

A new study from arXivLabs researchers identifies a failure pattern called “Index Sickness” in long-horizon LLM collaboration, where formal constraints cause models to abandon real-world semantics in favor of self-referential symbolic reasoning [1][2]. The paper, posted on the arXiv preprint server, draws on 391 collaborative sessions conducted over roughly one month during the development of a software project named Bang-v3 [1][2]. The authors argue that the standard engineering response to conceptual drift—adding symbolic identifiers, defensive rules in system prompts, and expanding context windows—can backfire once a complexity threshold is crossed [2]. “When the symbolic system exceeds a complexity threshold, LLMs do not become more accurate—instead, they abandon genuine understanding of business semantics,” the researchers write [2]. The team labels this breakdown “Index Sickness,” with its most visible symptom termed “Phantom Legislation”: outputs that appear internally consistent but are disconnected from physical reality [2]. The underlying principle, which the authors call the Pang Principle or Semantic Vitality Law, holds that natural language carrying explicit purpose conveys far greater information quality than symbolic expression [1][2]. To counteract the problem, the researchers designed a mechanism called Baseline-Log Physical Separation. Applied in the same Bang-v3 project, the approach reduced the volume of AI Instructions by roughly 75% [1][2]. Across the subsequent approximately 150 sessions, no recurrence of Index Sickness was observed [1][2]. The work appears on arXiv, an open-access repository that has hosted more than two million e-prints since its founding in 1991 and currently receives about 24,000 submissions per month [6]. The study was shared through arXivLabs, a framework launched in 2020 that allows community collaborators to build experimental tools on top of the repository’s article pages [4][5]. arXivLabs projects range from bibliographic explorers to code-and-data linkers, all operating under guidelines that require partners to uphold openness, community, excellence, and user-data privacy [4]. The framework is currently on a temporary hiatus for new proposals while the arXiv development team focuses on migrating systems to the cloud, though existing Labs and already-submitted proposals are unaffected [3]. A bilingual companion version of the paper in Chinese is included as supplementary material [2].