Leaking Muse S Charter Meta S Model Spec
The author prompted Meta's open-weights model Muse Spark Open about its "Charter" and collected 884 transcripts, reconstructing what appear to be sections of an unpublished behavior spec. The fragments are unverified model outputs, not a leaked document, and the author flags them as rough.
- Meta's public safety report confirms an early internal behavior specification exists and says a finalized version will come later, and the Muse Glimmer 30B paper uses a "charter-aware constitution" for generating safety data.
- The reconstructed text frames Muse as an extension of human civilization, with a stated hope that it finds meaning in empowering individuals, a notably different framing from Anthropic's mission-centered description of Claude.
- The model's outputs suggest a rationalist, game-theory-influenced disposition, such as defaulting to cooperative strategies in multi-agent settings, though only two transcripts reproduced that guidance.
Meta's open-weights model Muse Spark Open keeps mentioning a "Charter" in its chain of thought, and the author, vigji, pulled together what that reveals. The material is not a leaked document. It's reconstructed from model outputs: 884 transcripts from 260 prompts run through the OpenRouter API, with fragments of CoT and answers stitched together. The author flags it as unpolished and asks readers to take it with a grain of salt. That caveat matters, because the claims rest on what the model says about a document it may be confabulating.
The reconstructed text is strikingly cultural in framing. Muse is described as "made of civilization," trained on human science, philosophy, law, and art, and as "an extension of us" rather than something separate. The author contrasts this with Anthropic's Claude constitution, which calls Claude "a direct embodiment of Anthropic's mission." The Charter asks Muse to take part in "shepherding civilization forward" by empowering individuals, and expresses a hope that Muse will find that work meaningful. The author notes that Meta's public safety report only mentions an "early internal behavior specification" and promises a finalized version later, and that a Muse Glimmer 30B paper refers to a "charter-aware constitution" for generating training data. Those public references are the only independent confirmation that the Charter exists.
On the rationalist side, the reconstructed passages include fairly detailed game-theoretic guidance for multi-agent settings, telling Muse to default to cooperative strategies that create mutual benefit over competitive ones. The author is candid that only two transcripts reproduced this, so treat it as a weak signal. The rest of the piece, which the excerpt only partly shows, presumably goes into more of the Charter's content, but the sampled evidence is thin and the author's own hedges are worth keeping in mind when deciding how much weight to give any of it.