Entropic approved the use of the “ghost” document to train the Claude model

Recently, an internal document details the AI ​​model Cloud 4.5 Opus (Claude 4.5 Opus) is revealed to be “the soulChatbot refers. This document actually helps to form the character and how the model interacts with the users. Anthropic also confirmed that the said document actually existed and was used in the model learning process.

Richard Weiss, the person who discovered this document, explained on the LessWrong website how he was able to access a set of internal model documents by using a prompt to view cloud system commands. In one of these documents, there is a reference to “Soul Overview” has existed. Weiss then asked the model to reproduce this document, and the result was an approximately 11,000-word file that apparently defined Claude’s personality and behavioral framework.

Opus Anthropic Cloud 4.5 AI “Ghost” framework

This document is based on principles safety and commitment model to Produce healthy outputs and safe It focuses and constantly reminds Claude that “being useful to humans is one of the most important missions of the model” and that it should not go into areas that conflict with anthropic moral red lines. Such documents are usually used to establish the tone, ethics, limits of accountability and responsibility of language models.

More interestingly, Weiss claims to have requested the document from the cloud 10 times and each time exactly the same text was produced, which he says greatly increases the likelihood of the document being genuine. Several Reddit users were also able to retrieve similar sections of the same document from the cloud, indicating that the model likely has access to a copy of it in its internal data or training memory.

Amanda Askell, a philosopher and member of Anthropic’s technical team, confirmed in a post on the X social network that the model’s output is “based on a real document” that was used during the learning process. He also said that this document is still being revised and its full version will be published soon. The model doesn’t always reproduce the internal documents perfectly, but recent outputs have been “largely consistent with the original,” according to Skell.