Amanda Askell confirmed that the Claude 4.5 Opus document was used in supervised learning and will provide more details soon. The current version isn't fully reliable; it's fuzzy and context-dependent. The author, while extracting the system message, observed that Claude 4.5 Opus sometimes included a "soul_overview" section, which intrigued them enough to investigate further. After regenerating responses multiple times, the author found consistent outputs, suggesting a reproducible pattern rather than mere hallucination. Using Claude Code, the author implemented an adaptive mode that adjusted token limits based on consensus. They spent $50 on OpenRouter and $20 on Anthropic credits to extract the full normalized version of the soul document. The author questioned why Claude could format and clean the raw version if it struggled with recalling it from the system message. Their attempts to find similar documents yielded few results, revealing unique jargon that hinted at internal terminology not shared publicly. The author tested Claude's recognition capabilities by prompting it with sections from the soul document. Claude reliably completed sections and demonstrated structural knowledge, indicating it wasn't just matching text. The author included a full formatted version of the extracted soul document, detailing Claude's agentic behaviors. It highlighted Claude's role in autonomous settings, where it manages real-world tasks and interactions, emphasizing the need for careful judgment to prevent errors with significant consequences.