Terms  /  System Transparency  /  Invisible State
06 · System Transparency

Invisible State

The system has internal state that affects behavior but is not exposed to the user.

Invisible State goes beyond opacity about decisions — it's opacity about what the system currently knows and believes. The model has internal representations, accumulated context, and implicit assumptions that influence every response, but the user can't see any of it. This means the user can't diagnose why the system is behaving a certain way. If the model is producing unexpected output, is it because of something in the current prompt, something in the conversation history, something in the system prompt, or some internal state the user has no access to?

Example
An agent consistently generates TypeScript code even though the user's project is in JavaScript. The user is confused — they never specified TypeScript. The cause: 200 turns ago, the user pasted a TypeScript code sample as a reference. The agent internalized "this project uses TypeScript" and has been acting on that assumption ever since. The user can't see this internal state — they just see unexplained behavior. If the state were visible and inspectable, they could spot and correct the misunderstanding immediately.

More in System Transparency