What is Jev AI? Decisions, memories, and dreams in an AI village

What is the Jev model good at?

TypeSafe describes Jev as a System One model: a model designed for focused judgments that software can use directly. A developer supplies a situation and typed questions. Jev returns structured answers with probabilities, allowing the application to act on the result without first extracting a decision from a paragraph of generated text.

This fits tasks such as choosing among available actions, classifying a request, judging relevance, or evaluating a candidate against a defined rubric. The useful pattern is a clear question, meaningful context, and an answer the application knows how to handle. The village applies that pattern repeatedly as its inhabitants encounter different circumstances.

Three ways to ask Jev a question

• Choice selects from supplied alternatives and returns a probability distribution and confidence. In the village, choices can concern an allowed action or a relevant memory.
• Score evaluates against predefined, ordered levels in a rubric and also returns probabilities and confidence. This is useful when an application needs a graded assessment.
• Noul returns a probability for a yes-or-no judgment. The village uses this question type in its memory-retention review. A probability near the middle expresses uncertainty about the judgment; it does not automatically mean a medium amount of a trait.

Why an AI village is a useful demonstration

A single model answer is a snapshot. A persistent AI world simulation lets you watch what happens after a choice. Work uses time. Rest changes readiness for later activity. Supplies and shared places connect one inhabitant’s circumstances to another’s. The next decision takes place in a world affected by earlier actions.

Jevs Village gives these decisions a small, understandable setting. Six inhabitants have different roles, needs, and experiences. Mira works at the mill; Bram farms; Nia tends animals; Tomas hauls goods; Sela keeps the inn; and Old Père is the village elder. They share a world, but they do not all receive the same personal history.

That makes the village a useful use case for Jev: many focused judgments meet changing context and persistent consequences. Visitors can follow a person across several updates and compare their actions with the pressures around them. The interesting result is the sequence of decisions and what it leaves behind.

How Jev helps a villager decide what to do

Imagine Mira has work at the mill, is growing tired, and encounters a neighbour. Continuing work, resting, and interacting could compete for her attention. This is an illustrative situation; the real options depend on the current world and which actions its rules permit.

The simulation assembles Mira’s situation and enumerates legal combinations of actions, targets, and dialogue. Relevant remembered experience can accompany those choices. Jev evaluates the offered options. The simulation then checks the selected action against the world before applying its effects.

This gives the model a defined responsibility. It helps judge what makes sense in context, while the simulation controls what an action can actually change. Choosing work does not conjure missing supplies. Choosing an interaction does not make an unavailable neighbour appear.

Remembering before deciding

A persistent agent needs access to useful experience, but sending its entire history into every decision would make the context grow without limit. The village first retrieves candidate memories from an inhabitant’s own episodes. Their current situation, recent experiences, and important events help determine the candidates.

Jev can then select relevant memory evidence for the immediate choice. This is retrieval: deciding which part of the past is useful now. It differs from retention, which determines what stays available in personal recall over time. Both are judgment problems, but they answer different questions.

What does dreaming mean in Jevs Village?

In this AI village, dreaming is a name for sleep-time memory review. When an inhabitant enters sleep, recorded experiences can become eligible for review. Jev is asked how worthwhile an experience is for lasting recall, considering matters such as goals, commitments, relationships, novelty, verified outcomes, and future choices.

The review operates on existing episodes. It does not invent an adventure that secretly happened overnight or turn imagined events into world facts. The practical purpose is to help an inhabitant carry useful experience forward while allowing less relevant details to fade.

The application translates Jev’s judgments into retention decisions. Useful experiences can remain in lasting recall; low-value ones can leave personal recall; uncertain cases can wait for another review. Rules protect active obligations and evidence that still needs to remain available. A model judgment therefore works alongside explicit memory safeguards.

Forgetting personally does not erase the simulation’s history. The underlying record can remain available for consistency and audit even when an inhabitant no longer retrieves it as a memory. Later reviews can also reconsider retained experiences as circumstances change.

Review happens in bounded batches and can be deferred if a service is unavailable. A villager does not have to stay asleep while a backlog clears. This mechanism is a practical example of using Jev beyond immediate action selection: it helps assess which experiences may matter to later decisions. It does not establish human-like dreaming or consciousness.

How decisions become a village story

Jev provides structured judgments. The simulation validates actions and keeps world state. A separate language model turns committed events into readable narration. These responsibilities work together: a decision becomes an event, and the event can become a passage in the chronicle.

The narration gives visitors a way to follow village life, but a sentence alone cannot move somebody, create food, or change a relationship. Those effects belong to the simulation. Keeping the story grounded in recorded consequences lets the same world support both an observable system and an unfolding narrative.

What can we learn from the simulation?

The village demonstrates an application design: give a decision model clear options and relevant evidence, validate its output, and preserve the consequences. The same design questions arise in other agents and workflows: what context should a model see, what may it choose, and how should the application handle uncertainty?

Jev’s probabilities help express the model’s judgment. Confidence for Choice and Score reflects the shape of the answer distribution; it is not a guarantee that an answer is correct. Noul does not have the same separate confidence field. A permitted village action can still be unhelpful, and an overlooked memory can still matter.

Jevs Village is a working demonstration rather than a general performance benchmark. Its behaviour depends on the model, the supplied context, the world rules, and service availability. Not every animated movement represents a new AI request.

Watch Jev AI in action

Open the live village, select an inhabitant, and follow their public activity alongside the chronicle. Look at their needs and surroundings, then return after later updates to see what changed. Follow several decisions rather than judging the whole system from one line of dialogue.

The public observer shows selected outcomes and summaries. Private memory text and sleep-review judgments remain private. You can explore the consequences of agent decisions while these guides explain the mechanisms behind them.