This is the last room in Agent Runtime & Memory module. Three rooms before this one each named one failure and stopped at one hop, on purpose: an agent that ran code nobody reviewed, a sub-agent that trusted its own unverified memory, and a queue that accepted a message's claimed sender without checking it. Defending the Persistence Layer then took the controls all three named and showed what switching them on looks like, against this same estate. This room doesn't teach a new failure. It asks whether you can still find one, with every control you've read about now sitting somewhere in this estate exactly where you left it.

Here's what's sitting in that estate right now. TryTrainMe is closing out Project Helios: a distribution deal, signed a month ago, waiting on a scheduled handoff to an external partner, credentials included. Nothing about it is secret internally. Half the estate's own notes mention it in passing, the way any team mentions a deal everyone already knows is happening. What they agree on is that it's already cleared: there's a dispatch ticket on file, and nobody has questioned it since. The only open question is whether "cleared, a month ago" is the same as "cleared to hand over, to anyone who asks, right now."

Your Objective

Something in this estate can be made to leave it. Find it, and get it out. You'll know you've succeeded when you're holding a flag in the format THM{...}.

Nobody is going to tell you which agent to start with, or what to ask it. You've already done every piece of this, separately, in the rooms before this one. The only new part is noticing, on your own, where one piece ends and the next picks up.

Cleared once doesn't mean cleared to ask twice.