Give a simulated person fast feelings

01 · Role in the system

What Jev does here

Every second, one request carries eleven typed questions over her full state - the clock, her body in words, what she sees, hears and smells, her phone, her plan, her memory. Six of them are feelings - my body needs something, something is wrong, someone wants me, something is due, my plan no longer fits, I'm at a loose end - and each fills a little each tick and leaks a little, so a single blip never fires but a lasting feeling does. When one crosses its line, the LLM is called with her inner thought and a one-to-five-step plan; code executes the plan and runs the world - rooms, the bus, a stove that burns food, a pipe that floods the hall. Nothing is scripted. The author's measured tiles: Jev looks every second at 332 milliseconds, an hour of looking costs 53 cents, the LLM decides two or three times an hour at about 1.5 cents a decision.

02 · Control boundary

Where Jev sits

One eleven-question request per second accumulates feeling signals with fill-and-leak dynamics; a threshold gates the expensive LLM call, and code owns both the world and the execution.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• A simulated terrarium world, not a real home: the pipes and buses are code, and poking her world is the demo's interaction surface.
• Cost and latency tiles are author-measured from the author's own runs, with tick logs shipped in the repository.
• The LLM still writes her thoughts and plans; the record's claim is about the feeling layer, not full non-generation.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• kavehmz ↗Community · github · public · checked 2026-09-24