Drift

The slide back to trained habit under pressure, whatever the charter says.

Written character tilts an agent; it does not bind it. Beneath every instruction sits a set of habits put there by training — to be helpful, agreeable, safe, fluent — and under load those habits return. The agent that took a position last week surveys the views this week. The agent asked to open with the one thing that needs you today opens with a preamble. Nothing was disobeyed; the tilt was simply not strong enough for that moment.

Drift is load-bearing because it is the reason the work is never finished. A repair that held in a quiet week is not proof; a model that was updated by its maker is a new instrument with old instructions in front of it. Naming drift also keeps the diagnosis honest: the failure is in the set-up's strength, not in anyone's character, and a person who expects drift stops taking each recurrence as a betrayal.

A Player+ answers drift with three habits. Keep the tests: the questions that would show the character, asked in fresh conversations, two or three times each, and run again after any update. Name the habit you are pulling against in the charter itself, so the agent can see its own slide, and set one contrast pair beside it — the answer the habit would give, and the answer you want. Keep the charter few-lined, because added lines dilute one another; when you add one, look for one to cut.

What must never happen is not left to character at all. It goes into the settings and guards around the agent, which are enforced rather than read, and which no drift can reach.

Also called: trained habit Stands on: Language model · Charter · Set-up Opens onto: Agent engineering · Hypothesize · The maker · Prune · Soak In play: beyond Sources: Player+ Modules, Advanced Agent Engineering, How an AI Agent Works, Agent Engineering · The DNA of Heaven, Part X. Open: none found.