twinkling-reality/exulanica

Worlds for AI Agents.
Website ·
Documentation ·
Roadmap ·
Getting started
Exulanica is Worlds for AI Agents. People create places and inhabitants, shape their
interactions, and develop worlds into projects they can explore and build on. Open models can
take the roles the engine registers. Comparing two models on the same saved world is one
capability.
Product direction owns scope, the first milestone, and the
hypothesis about discovering models through a world.
Hosted model calls go to Nebius Token Factory. The
model manifest names the models, including NVIDIA
Nemotron, Qwen, and DeepSeek. Exulanica does not generate a world's appearance with a learned
model.
Product direction. The capability boundaries below distinguish implemented foundations from delivery targets.
Models inside the world
A person builds a world. The things in it that act are its agents: people, animals, vehicles,
and decision-makers such as a shopkeeper. Open models can take the roles the engine registers.
Comparing two models on the same saved world is one capability. Exulanica does not generate a
world's appearance with a learned model.
A model is a proposer, not an authority. Two decision roles are registered: a person, and a
generated town's junction signal. Each role's contract declares what the model is shown and which
actions it may choose. The engine validates every choice before it takes effect, and every accepted
decision is stored, so a run replays without calling the model again and two runs from the same
saved world compare fairly. A junction signal may keep a green light one second longer or switch
to amber within its plan. No evaluation record measures whether that improves traffic, and no
improvement is claimed. Vehicles, animals, weather, and shopkeepers are not model roles. Every
hosted model call passes one policy boundary that enforces the egress allowlist, the budget, and
the person's rights.
Several open models serve hosted roles through that boundary (the
model manifest). People in a starter or a world made
from photographs follow a deterministic planner until the owner chooses a model. A generated town
is a separate saved world. What the first milestone accepts, including generated towns and what
the Compare view runs, is the status table in
product direction. Not built: roles beyond a person
and a junction signal; vehicles, weather, or an economy run by models; and a retraining loop.
The world
The experience has six connected parts; product direction sets the
order they are delivered in:
Models run your world
Open models decide what the people in your world do: where to go, whether to wait, stand a while or stop to talk. The engine validates every decision and stores it, so a run replays exactly. The world's owner chooses the model for each person or group of people; with no choice, their routine decides.
People and models · Decision roles · Model selection
Swap a model, see the difference
Run the same saved world's hour under two models and compare what happened. A comparison scores how the people fared against a control run and shows both runs side by side in the application, where the world's owner starts it within a spending bound they state and watches its runs finish.
Controlled comparisons · First milestone
Build your world
Place pieces from the catalogs, make a town from a served recipe, make a world from photographs, or import content with its origin labelled. Edits persist as versions. A starter and a world made from photographs each start with eight people. A generated town's people come from its homes.
World creation · Reconstruction
Give the world life
The people in your world follow routines, walk to the places they use, and respond to what you change. Small birds fly and perch on the trees that host them, on a clock everyone watching shares. Pause and playback speed are clock controls. The world clock contract owns the rest of the clock. Calendar, weather, and fictional physics are not configurable.