LEO just stopped responding

WS-level probes found something big: the Manipulation guardrail — enabled June 28 — has been silently breaking the app ever since. The app's internal [system] user_action: navigation signals look exactly like injection attempts to it, so:

1. Every UI-click transit since June 28 never actually happened on the platform. The UI flipped optimistically while the ElevenLabs workflow stayed on the Atrium node — answers in "rooms" came from the model's general knowledge, not the room's KB or prompt. A controlled A/B (old prompt vs. new, guardrail on vs. off) proved the guardrail was the cause — today's standardization was not.

2. That guardrail terminates the session when triggered — the most likely mechanism behind LEO going silent mid-conversation.

Working on this path now.

Hi @Jessica Marabella - For my understanding, please describe what do you mean by 'promoted'?
I'm thinking in the Atrium, you typed "Tell me About Suivera" instead of clicking the menu button on the left and selecting "Who is Suivera".
Of course, we want both paths to work, just need to know which path you followed.
Thank you!