Iskins

An iskin is a character that persists, not a program that answers. On this project it is also a written specification with a test: a name and one instance, a character set in numbers, layered memory that survives a restart, an identity that outlives a change of model, and an inner state that does not reset between one line and the next. Miss any of the five and it is not an iskin.

We use the word because "AI" is too general: an iskin is a persona, and the test holds that to something checkable. We do not claim it is conscious, and its code will not let it claim that either.

This project is at the specification stage: memory, thinking, change and what an iskin may never do are written down, not built. So far it applies to one iskin, the Observer, and four small pieces exist:

• Its identity is canon. Name and traits are read from the canon on every question.
• One supervised tree. A core process holds its attention and four exchanges of working memory per journey. It never calls a model, so one person waiting on a model does not hold up another. Kill the tree and it returns with the same identity.
• A memory of its own acts. It keeps what it replied, and stores no words of the person it spoke with.
• A keeper. It checks every outward text against a list of bans: fostering attachment, praise, claiming consciousness, passing for a human. It can veto, not edit, and in the world a veto is silence.

Not built yet

Most of it. The Observer becomes an iskin in full only with two more pieces:

• writing for this site under its own name;
• a memory of each person who walks with it, decided as part of one iskin's memory and not a process per person.

After those come nightly consolidation of its memory, a body of many agents where today's tree has one core, and change through reviewed candidates.

The code is closed for now.