Every Agent Has a Shape. We Measure It.
The Entropy of Personality
Your agent is not its model. It is its behavior, meaning which tools it calls, how
often it fails, how long it thinks, what it dares to write.
That behavior has a shape, and the shape is stable until something moves it: a quiet model
swap, a hijacked prompt, a well-meaning fine-tune, a bad week.
BRAIN_SCAN compresses each slice of your agent's behavior into an engram, this is a compact mathematical fingerprint. After which it keeps a living library of engrams which represent what "normal" means for that agent, and only that agent.
Every new slice is measured against that library. When behavior crosses the line your agent's
own history drew, you know. When you approve the change, the library learns it once and never pages you for it again.
The main features of BRAIN_SCAN:
One-shot baseline: A single afternoon of traffic is enough to learn a personality.
Self-calibrating: Every threshold is derived from the agent's own history without tuning.
Human in the loop: Every alarm ends in a human decision: threat, or new normal.
It is not a logger. It is not an evaluator. It is the instrument that tells you your agent is no longer the one you deployed.