If you lock part of an AI behind a password, what does someone have to do by hand, once, to set it up?

This explores what up-front human effort it takes to set up a hidden, locked-away region of a model or its context, as opposed to costs that keep recurring afterward. The corpus has no note that itemizes those costs. The closest material is one note on password-locked context capsules, and it only hints at the answer.

That note reports that the paper's experiments used locks that had already been installed, with passwords the researchers knew, which made building a capsule straightforward Can you build a capsule without knowing the password?. So the one-time human cost visible here is small and specific. Someone has to decide there is a lock, install it, and hold the password when the capsule is built. The note says it is unclear whether a capsule could be built without the password or found by other means. Any cost of guessing, recovering or discovering a hidden partition is therefore unmeasured. The experiments show only the cheap case, where the person building the partition also controls the key.

Another note suggests the setup cost is only part of the bill. It describes treating distilled expertise as versioned files that can be inspected, corrected and rolled back, so nothing important sits in hidden prompt state Can person-grounded skills remain auditable without hidden prompt state?. The implication is that hiding something removes the audit trail. If a partition is hidden, keeping it trustworthy means paying for the inspection, correction and rollback that the file-based approach gives you by default. That looks like a recurring cost rather than a one-time one.

One more framing helps with 'one-time'. A 115-day case study of persistent agents argues that when context is reused, cost is better counted per finished artifact than per token Do persistent agents really cost less per token?. A one-time cost, such as setting up a lock, is spread across every later use of what it protects. That is a general accounting idea and not a finding about hidden partitions. If you want a real answer, the corpus would need a note that measures the human setup work, such as choosing what to hide, writing the lock, and testing that it holds.