August 19, 2026

[cassidyjames]

Attendees: @annebdh, @julietshen, @cassidyjames, @jeremieponak-dot, Alice, @ThatKoffe, @ThisIsMissEm, Nav

• Roblox models joining RMC
  • PII classifier: real-time filtering for asking for or giving PII (usernames, phone numbers, addresses, etc.)
  • Sentinel
  • Voice safety
• roostorg#82
  • How can models “speak a common language” or use standard formats?
  • Musubi is building agent skills for formatting policies
  • Policies are written in plain language markdown and then tested against specific models
    • A lot of A/B testing to find difference between policies and/or models; it’s easy to see when there are differences, but hard to know why.
• @julietshen has a side project called vibecheck for testing/evals
  • Tested Roblox Sentinel
  • Compared steerability using inverted policies with interesting results: Shieldstral was not steerable in this way, CoPE-A and CoPE-B were
  • Could be interesting to reach out to platforms designed for or with a large amount of consensual adult content to develop synthetic datasets based on their content to use with these types of evals