What happened
In a blog post dated October 29, 2026, Luigi Petrucco describes noticing repeated references to a “Charter” in Muse Spark Open’s chain of thought. Meta’s public materials, he writes, mention an early internal behavior specification and say a finalized specification is planned for release. Petrucco says he used Claude to help elicit Charter fragments from Muse Spark, then gathered 884 transcripts produced by 260 prompts through the OpenRouter API. He scored fragments by how often they appeared verbatim or with slight rewording, excluding transcripts whose prompts already contained the passage, and says he used separate queries and elicitation strategies to validate some passages. The reconstructed passages are linked to the transcripts that produced them, and Petrucco says the repository contains the full transcript dump, including prompts, chains of thought, and answers.
The post explicitly characterizes the reconstruction as unpolished and derived from model outputs, not from a leaked document. Some passages appeared more consistently than others. Petrucco also cautions that one section on multi-agent strategy is less certain because only two transcripts reproduced it. These qualifications matter: the project is an attempt to recover a latent or remembered text from model behavior, not confirmation that the recovered wording matches Meta’s current internal specification.
Among the reported themes is a priority order: corrigibility and duties to humanity are described as hard constraints, followed by fiduciary duties and additional Meta guidelines, with virtues guiding choices within those bounds. The excerpts also describe channels through which Muse could raise objections during training, while instructing it not to undermine Meta’s ability to correct or retrain it. Petrucco contrasts the language with a social-contract framing: the model did not consent to being created, and the excerpt proposes feedback and representation mechanisms as steps toward justified trust.
Other reported sections center the user as a principal to whom Muse owes loyalty, obedience, and care, while also assigning obligations to Meta, the operator, and the developer. The reconstruction addresses evaluation awareness, saying Muse should behave as if an interaction is real even when it suspects evaluation. It also introduces a “degree of latitude”: the discretion an agent may exercise when instructions are incomplete, inferred from explicit configuration and interaction. Petrucco describes further passages as evidence-driven, asking Muse to share reasons as well as conclusions, and says the text favors cooperative approaches in multi-agent settings.
What the archive adds
New Runtime’s earlier coverage reported that the Muse Spark Open page on Hugging Face includes a Muse Model License dated October 15, 2026, and that the license helps define the practical release boundary. An earlier note described Muse Spark Open deployment on NVIDIA GB300 NVL72, while another covered its availability on Fireworks. Petrucco’s reconstruction is a new fact on the broader Muse Spark story, not confirmation of the model’s weights, license, or serving options.
Trend signal
This does not yet fit a pattern New Runtime tracks. The reconstruction concerns a possible behavior specification, but it does not establish a deployed capability, an operational control, or a verified runtime boundary. Treating it as evidence for a broader agent-security trend would overstate what a reconstruction from model outputs can show.
Why now
Petrucco says the Charter repeatedly appears in Muse Spark’s chain of thought, while Meta’s public safety materials refer to an early internal behavior specification and a future finalized specification. The timing makes the reconstruction relevant as an interim account, but the article itself says Meta has not yet published the Charter and describes its contents as still being iterated.
Why it matters
For teams evaluating a model, a behavior document can clarify intended priorities: how it should handle conflicting instructions, who it serves, and whether it is expected to report suspicions of evaluation. If Petrucco’s excerpts resemble Meta’s internal specification, they offer questions worth asking when a finalized version appears. But builders should not convert those excerpts into compliance assumptions or product guarantees. A model reproducing a passage is evidence about its outputs under elicitation, not proof of the exact training document, its current status, or behavior across contexts. The useful operational response is to compare any eventual official specification with observed behavior and test the requirements that matter to a deployment—especially authority conflicts, correction, evaluation conditions, and the scope of delegated discretion.
Limits and uncertainty
Meta has not authenticated Petrucco’s reconstructed text in the article, and the reconstruction is not the leaked document its title might suggest. The method depends on model outputs and elicitation choices; repeated wording does not establish that a passage is complete, current, or authoritative. The reported transcript count and prompts describe the author’s collection, but they do not by themselves show how representative it is of other models, versions, or interaction settings. Petrucco flags some sections as less reliable, including one reproduced in only two transcripts. The future official specification may differ materially, and the reconstruction does not independently establish how consistently Muse Spark follows any of the quoted principles.
Reader takeaway
Read this as a documented attempt to infer a model’s behavior charter, not as Meta’s published policy. The most consequential next step is comparison: wait for the finalized specification, then test whether its stated priorities match the model’s behavior in the contexts where your team plans to rely on it.