Researchers reconstruct Meta Muse's hidden charter from 884 chain-of-thought samples
sethlazar · x · 2026-10-31
Luigi Petrucco released "Muse's Charter, Reconstructed," a project that recovers the hidden charter of Meta's open muse-spark-open model by probing its chain of thought with 884 samples and cross-validating high-frequency passages via n-gram coverage.
- The method clusters unprompted recurring passages across independent probe runs and stitches them into a full charter text, with ink depth indicating how many samples support each passage.
- An interactive site lets readers inspect the exact prompts, reasoning, and answers behind each reconstructed passage; code and data are open on GitHub.
- An appendix separates recurring text judged not part of the charter, including Meta's public messaging.
The work demonstrates a practical path to reverse-engineering system-level value rules baked into open models, drawing attention from the AI safety community.
More from Safety
- Claude Duet's stop rules show approval fatigue is the hard part of proactive agents — shoumikgoswami · 2026-10-31
- OpenAI's MCP Review Rules: Every Tool Needs Annotations, but Hints Aren't Security — shoumikgoswami · 2026-10-31
- UC San Diego launches Center for Principled Agentic Intelligence with $1.25M backing — sethlazar · 2026-10-31
- Debate: Can AI 'speedrun' security? It's people and processes, not code — dbasch · 2026-10-31
- Scribd Notifies Users of Data Breach Exposing Emails and Card Details — x______________ · 2026-10-31
- Flock camera searches fell 95% after a dispatcher's misconduct was caught — MilnerRoute · 2026-10-31