Small decisions. Interesting possibilities.Submit contentSubmit
Jev Town GIF preview

Jev Town

Simulate characters whose structured actions come from Jev and conversations come from a language model.

Added to Jevfast

How it uses Jev

JevTown has an optional action decider that sends an agent context as named state fields and several typed questions in one System One request. It asks Noul questions about whether the agent should seek a target or roam, then uses Choice questions to select a target, place, and idle duration. Application code composes these answers into an approach, wander, or idle action and maps chosen labels back to manifest IDs.

What Jev decides

Selected questions from an illustrative JevTown movement requestExample answers · not a recorded Jev response · Source ↗
Question 1 · seek
YOUR APP
INSTRUCTION

Decide whether it is a moment to go to somebody or something rather than stay where Alice is.

STATE

Source-owned fixture: Alice is a cautious millwright. Her state says the mill runs hot and she means to ask, and the world rule says nobody can leave town. The example manifest contains Bob, the Mill door, and the mill yard. This chart shows a subset of the request questions; the destination Choice is omitted.

JEV · NOUL
YesNo
Question 2 · target
YOUR APP
INSTRUCTION

Choose which available person or object is worth going to now.

STATE

Source-owned fixture: Alice is a cautious millwright. Her state says the mill runs hot and she means to ask, and the world rule says nobody can leave town. The example manifest contains Bob, the Mill door, and the mill yard. This chart shows a subset of the request questions; the destination Choice is omitted.

JEV · CHOICE
  1. talk to Bob
  2. go to Mill door
Question 3 · roam
YOUR APP
INSTRUCTION

If not approaching somebody or something, decide whether walking somewhere else is better than staying put.

STATE

Source-owned fixture: Alice is a cautious millwright. Her state says the mill runs hot and she means to ask, and the world rule says nobody can leave town. The example manifest contains Bob, the Mill door, and the mill yard. This chart shows a subset of the request questions; the destination Choice is omitted.

JEV · NOUL
YesNo
Question 4 · idle length
YOUR APP
INSTRUCTION

If staying put, choose how long before it is worth looking around again.

STATE

Source-owned fixture: Alice is a cautious millwright. Her state says the mill runs hot and she means to ask, and the world rule says nobody can leave town. The example manifest contains Bob, the Mill door, and the mill yard. This chart shows a subset of the request questions; the destination Choice is omitted.

JEV · CHOICE
  1. pause for a moment
  2. stay put for a while

App workflow

  1. Build agent context and manifest

    The app combines the active agent description and state with world rules, then gathers the current manifest of available targets and places.

  2. Create typed questions

    The Jev decider creates Noul gates and Choice questions. It phrases target and place choices as readable actions and keeps internal IDs in a separate mapping.

  3. Call System One

    The operations layer sends the generated state and questions to the separate System One provider and records the request and typed answers in the trace.

  4. Compose a world action

    The decider interprets the Noul probabilities and Choice answers, applies local thresholds and idle policy, then maps the selected label back to a manifest target or place.

  5. Apply the action in the demo world

    The operation sends an approach, wander, or idle input to the agent simulation. The Solarium demo renders the agents and their decision stream.

Why it is interesting

The decider makes the action set explicit before asking Jev. Choice labels describe the moves in ordinary language, while the app keeps internal IDs out of the model-facing options and maps the returned label back to the world manifest. A Solarium demo uses the same agent loop on a scene-backed world.