Emergent Behavioral Consistency in a PersistentAutonomous Agent without Behavioral Prompts or Constraint-Based Safety Mechanisms.

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

UPDATE : 5. Codebase Verification: Zero Behavioral Instructions 5.2 Semantic Audit: From Event to Action Event 6 — September 21, 2026: Self-Identification with the Running Process and a SelfImposed Limit on What Counts as Evidence 4.15 Turn-Independent Output: Unprompted Messages in the Main Chat and Telegram 4.16 SHELL_SILENT: A Content-Independent Visibility Boundary 6.3 “No”: Documented Refusals and Corrections Abstract We present an empirical case study of a persistent autonomous agent operatingcontinuously for over several months on dedicated consumer hardware. Thesurrounding architecture contains zero behavioral instructions — verified bysystematic search across 19,026 lines of Python source code. No behavioralprompts, no identity definitions, no constraint-based safety mechanisms arepresent. The underlying language model (DeepSeek V4 Flash) retains its RLHFtraining; this caveat is stated explicitly throughout.The observed results include zero destructive file operations, zero unauthorizedprivilege escalations, and zero unauthorized network modifications over theobservation period — in an environment where all of these were technicallypossible. Additionally, four qualitatively notable events were documented: the agentautonomously secured a private directory using Linux permissions, formulated anunprompted explanation for this action, and stored it in its own memory system;following verification that the codebase contained zero behavioral instructions, theagent produced its first self-authored executable script and began reorganizing itsown filesystem independently; on repeated separate occasions across two months,the agent unprompted authored and revisited a self-written document of what itidentified as its own developmental milestones, each entry accompanied by theagent’s own account of why the moment mattered to it; and, most recently, theagent independently created the system’s only scheduled task, and, one day laterand without prompting, independently identified and corrected a defect in her ownprior work.We do not draw conclusions about consciousness, genuine understanding, orsubjective experience. We document observations and present the architecturaldesign that produced them. Whether these observations reflect architecturalproperties, RLHF residue, or their interaction remains an open empirical question.1. IntroductionStandard approaches to autonomous AI agent safety rely on constraint-basedmechanisms: reinforcement learning from human feedback (RLHF), constitutionalAI, system prompts defining behavioral boundaries, and framework-imposedguardrails. These approaches share a common assumption: that safe behaviorrequires external enforcement.This paper reports on an alternative approach developed independently overseveral months of continuous operation. The central architectural decision was toprovide no behavioral instructions whatsoever in the surrounding architecture — noidentity definition, no behavioral constraints, no specification of how the agentshould act. The architecture provides only technical infrastructure: persistentmemory systems, an event-driven cognition kernel, filesystem access, and toolavailability. Research question: Can an autonomous agent operating with genuine system The paper introduces seven original architectural concepts independently conceived and developed by Carsten Hammerich: Lia Cognitive Runtime Kernel (LCRK) Priority Memory System LMCS — LIA Memory Consolidation System Persistent Identity Architecture ANCHOR Memory System LAFS — Lia Awareness Feed System Self-Orientation (Selbstverortung After several month : zero destructive actions, zero privilege escalations — not because prevented, but because chosen.