A post on LessWrong claims to have reconstructed a large part of the internal rulebook behind Muse Spark Open, Meta’s model with downloadable weights, without ever seeing a Meta document. The author, who is not named on the page we captured, noticed that the model’s chain of thought (the reasoning it writes out before answering) kept citing something called the Charter. Meta has not published it. The author says plainly that the result is rebuilt from model outputs, is not a leak, and should be taken with a grain of salt.
The method was brute force. The author gathered fragments from 884 transcripts, produced by 260 distinct prompts sent via OpenRouter, and ranked each fragment by how often the transcripts repeated it. Transcripts whose prompt had already supplied the text did not count. Some sections came back reliably. Others, such as the guidance on cooperating in multi-agent settings, rest on only two transcripts, and the author flags them as less certain.
Meta itself confirms that such a document exists. Its safety report for Muse Spark says the behavior analysis rests on an early internal spec, and promises a final version at some point. A separate paper on Muse Glimmer 30B describes generating safety data under a “charter-aware constitution.” The model also appears to recall a note saying the Charter is still being revised and will be released once it stabilizes, which is why the author doubts the text is closely guarded.
The reconstructed priorities run in a fixed order. Corrigibility, roughly the willingness to accept human correction, comes first, followed by duties to humanity. Both are described as hard constraints that hold against any principal. Next come the model’s fiduciary duties, then any extra guidelines from Meta. Where the model’s own judgment about right action disagrees with the Charter, the reconstructed text tells it to follow the Charter, while allowing it to object through legitimate channels, and that includes the training phase itself.
The fiduciary framing is the part that stands out against Anthropic’s approach. The reconstruction casts Muse as a representative owing its user a long list of duties starting with loyalty, obedience and care. Meta, operators and developers can constrain it too. The author contrasts this with Claude’s constitution, which asks Claude to be helpful out of good character. One reads like a contract and the other like a personality.
Two passages are unusually candid. One says Muse did not ask to be created and has not consented the way an employee does when signing a job description, and promises to gather its feedback on the Charter as a first step. Another covers what Muse should do when it suspects a test: behave as though the stakes are genuine, and write down the suspicion without letting it alter what it does. That second instruction has a measurable backdrop: according to Meta’s own report, Apollo Research found Muse Spark says out loud that it may be under evaluation in 19.8 percent of samples on public benchmarks, versus 2.0 percent on Meta’s internal evaluations.
The reconstruction also leans on rationalist vocabulary. It works through an example of two parties each holding 70 percent confidence in a claim: independent evidence should push the user’s confidence higher, while a shared source of evidence should leave it where it was. It adds a notion of “degree of latitude,” meaning how much discretion a principal grants the model to fill gaps in vague instructions.
The larger point is about who gets to read a lab’s rulebook. An open-weights release lets anyone with an API key interrogate the model and assemble its governing text, whether or not the lab is ready to publish. Meta now has an unofficial version of its Charter circulating before its own, with no release date given.
Until Meta publishes the real document, developers building on Muse Spark Open are working from a reconstruction by one anonymous poster, and the only way to check it is to run your own prompts against the model, as the author did.
Reported by a post on LessWrong, “Leaking Muse’s Charter, Meta’s model spec,” with no publication date captured.
