Renaissance fresco of scholars observing a miniature inhabited valley
Wega Labs

Wega Labs · Persistent agent research

A little world that keeps living.

Eighteen artificial residents begin in a valley. They have private memories, relationships, needs, and commitments. The world continues to update while no observer is present.

Current phasePhase 1 · deterministic kernel and replay

How December is built.

Formal model, state integrity, agent decisions, identity experiments, validation, and current implementation.

Read the paper

System objective / 01

A persistent world with observable causes.

The simulation runs without continuous human input. Residents can work, communicate, form relationships, create institutions, migrate, reproduce, and die.

When an observer returns, the system reports what changed, which events caused it, and what information each resident had when they acted.

Renaissance scientific plate showing a person connected to memory, relationships, choices, environment, and stages of life
Plate II Components of an artificial life history

Research subject / 02

Identity across time and change.

December studies whether an agent remains behaviorally distinct as its body state, memory, relationships, environment, and underlying model change.

  • Observation

    Information available to the resident at a specific time.

  • Memory

    Sourced experiences, testimony, beliefs, commitments, and retrieval errors.

  • Relationships

    Shared history, obligations, trust evidence, and social recognition.

  • Decision

    A selected command under material, social, and informational constraints.

Initial scenario / 03

A bounded population in a causal environment.

01

Physical state

Weather, water, food, materials, distance, work, health, and time are explicit state variables.

02

Individual agents

Each resident has private observations, memory, relationships, needs, commitments, and available actions.

03

Causal history

Every accepted action produces events that record causes, authorization, state changes, and random draws.

Overhead illustration of the Founding Valley with eighteen warm points connected by causal traces

System architecture / 04

Language models propose actions. The kernel updates the world.

01Observation

Visible state, body state, accessible records, and delivered messages

02Memory

Sourced episodes, beliefs, relationships, commitments, and procedures

03Typed command

Schema, actor, intent, targets, parameters, and believed preconditions

04Authoritative kernel

Validation · state transition · scheduler · named randomness

05Event history

Immutable events, causal parents, authorization, hashes, and replay

06Observer

State projections, timelines, causal graphs, metrics, and branches

Only validated commands can change canonical state. Memory, narration, model output, and the observer interface have no direct write access.

Model constraints / 05

Seven requirements for a valid run.

Each requirement maps to state variables, transition rules, observable evidence, and tests.

01

Material conservation

Food, materials, tools, time, and labor cannot be created by model output.

02

Time and location

A resident occupies one location. Travel, work, care, and communication require time.

03

Information boundaries

Agents act on observations, testimony, beliefs, and memory rather than hidden world state.

04

Heterogeneous objectives

Residents can assign different weights to safety, kinship, status, duty, reciprocity, and other goals.

05

Causal structure

Aggregate outcomes must be derivable from prior physical processes, decisions, and interactions.

06

Persistent state

Injuries, resources, obligations, institutions, relationships, and ecological changes persist across model calls.

07

Claim limits

Simulation output does not establish human behavior, consciousness, or personal survival.

Observer / 06

Inspect state, events, causes, and counterfactuals.

01

State change

Population, stores, structures, health, territory, institutions, and ecology.

02

Event history

Births, deaths, work, decisions, disputes, projects, migration, and hazards.

03

Causal trace

Paths from an outcome to prior decisions, observations, constraints, and random draws.

04

Counterfactual branch

A separate simulation starting from a selected decision or intervention.

01Rainfall
02Stream flow
03Available labor
04Harvest
05Stores
06Bargaining power
07Health + migration

Validation / 07

A single run is not evidence.

01

Four-resident pilot

Test continuity measures and controls before connecting cognition to the full settlement.

02

Kernel ensembles

Run many no-LLM worlds to locate invariant failures, unstable parameters, and trivial outcomes.

03

Ablations

Remove memory, relationships, scarcity, institutions, or migration and measure the change in predefined outcomes.

04

Replay and branching

Reconstruct recorded history exactly. Treat changed decisions or new model calls as separate branches.

05

Complete reporting

Publish positive, negative, null, and unstable results with the same protocol and artifact requirements.

Launch acceptance criteria

0

Known invariant violations in stress ensembles

100%

Replay equivalence for golden histories

Bounded

Provider failure recovery and compute use

Traceable

Causal path for every reported outcome

7 days

Running unattended before the first public world wakes

Research boundary / 08

December studies operational continuity, not consciousness.

The current experiments measure continuity in commitments, relationships, beliefs, adaptation, and behavior across time and perturbations.

A positive result does not establish subjective experience, personhood, or survival of a human or artificial subject.

Open research / 09

Specifications, code, experiments, and results are inspectable.

01

Protocols, parameters, assumptions, and outcome definitions are versioned before confirmatory runs.

02

The kernel, event model, experiment harnesses, and analysis code are open source.

03

Recorded runs can be reconstructed and inspected at the event level.

04

Negative results, model refusals, costs, deviations, and known invalid regions are published.

Claim limit

December is a research system for persistent synthetic agents. It is not a validated model of humanity or a test of consciousness.

Contribute / 10

Review or contribute to December.

Inspect the specification, challenge an assumption, contribute code, or propose a test.

B

Propose a test case

Suggest a decision, relationship, institution, hazard, or ordinary activity that the model should represent and explain.

December currently uses fictional residents only. Please do not submit anyone’s private life history or nominate a human participant.

Open an issue ↗