Managed Memory Is Still Not Governed State: Production Lessons from LLM Wiki v2

This release presents Article 31 (P31) from the Reflexive Laboratory research program, a transcript-based approach to AI-assisted scientific inquiry. The paper examines a recent extension of Andrej Karpathy’s LLM Wiki idea, often described as a “compiled memory” system in which documents are incrementally transformed into a persistent, interlinked knowledge base. The extended version analyzed here introduces practical mechanisms developed through real-world use, including confidence scoring, supersession of outdated claims, knowledge graphs, hybrid search, automation hooks, audit trails, and collaboration support. Rather than treating these additions as a replacement for the original idea, the paper interprets them as evidence of what happens when compiled-memory systems are used at scale. As such systems grow, they begin to require lifecycle management, structural organization, and accountability mechanisms to remain useful over time. The central argument is that these developments produce a stronger form of managed memory, but still do not yield the more demanding object defined in the Reflexive Laboratory series: a governed, state-bearing research system in which authority, validation, and current working state are explicitly represented. This package includes: the final publication-ready manuscript (PDF), the source manuscript, a structured comparison table, supporting materials used in the analysis, and the transcript record documenting the construction process. The goal of the release is not only to present a comparative result, but to demonstrate a research workflow in which both the final paper and its production trace are preserved as part of a unified research object.