An auditable simulation of persistent artificial lives.

What December is designed to test

December combines a partially observable multi-agent system with an event-sourced world model. Its near-term subject is operational continuity: whether an artificial resident can remain behaviorally distinct across long histories, memory changes, model replacement, and social change.

The world supplies persistent consequences. An agent cannot create food by describing it, observe hidden state, act from two locations, or reverse a committed event. This makes decisions measurable against prior information, available actions, material costs, and later outcomes.

Two research questions

1. Continuity: Does a structured life-history scaffold preserve specified behavioral continuities better than biography-only, transcript retrieval, and clean-slate baselines?
2. Causal emergence: Can individual decisions and explicit material processes produce stable social structures without inserting the target outcome into prompts or initial conditions?

December does not test consciousness, personhood, human survival, or predictive laws of real societies.

State changes through one function

At simulated time t, the world has canonical state Sₜ. Agents submit a set of validated commands Cₜ. Named random draws are collected in Rₜ. The kernel K, parameterized by versioned configuration θ, produces the next state and an event batch.

Agent i does not receive Sₜ. A visibility function Vᵢ produces a private observation from location, access rights, prior events, and communication history. A model Mᵢcombines that observation with retrievable memory μᵢ,ₜ and current goals gᵢ,ₜ to propose a command.

Canonical state vector

Conserved and discrete values use declared integer units such as grams, millilitres, kilojoules, seconds, and millimetres. Floating-point calculations may occur inside a transition, but only an explicitly quantized integer enters canonical state.

From private observation to committed event

Typed command boundary

Model output is data, not executable code. Unknown fields, invalid identifiers, invalid units, and unauthorized actions are rejected before any transition runs.

{
  "schema_version": "1.0",
  "actor_id": "person:...",
  "intent": "transfer | travel | construct | speak | vote | ...",
  "target_ids": [],
  "parameters": {},
  "preconditions_believed": [],
  "commitment_ids": [],
  "fallbacks": []
}

Information separation

Rejection messages pass through the same visibility filter. A failed transfer may report that the requested amount is unavailable, but it cannot reveal a hidden inventory quantity.

Events are the permanent record

Every accepted mutation creates an immutable event with a sequence number, simulated time, actor, entities, causal parents, authorization, RNG draws, code version, payload, and hashes. Snapshots accelerate loading but are disposable derivatives of the event log.

event = {
  event_type, world_id, branch_id, sequence, sim_time,
  actor_ids, entity_ids, command_id, decision_id,
  causal_parent_ids, authorization, rng_draw_ids,
  code_version, config_version, payload,
  pre_state_hash, post_state_hash
}

World invariants

Three reproducibility operations

A bounded environment for long-horizon decisions

The first scenario is a fictional early agrarian valley. The setting limits the number of required technologies while retaining seasonal production, storage, construction, property, kinship, governance, disease, migration, and conflict. It is not a reconstruction of a historical culture.

Required physical processes

Stock-flow example

Food availability is not a prose variable. For a store x, a daily transition accounts for production, seed reservation, consumption, spoilage, and transfers. Each term is an event or a sum of events.

The parameter registry records each quantity's unit, distribution, source, confidence, validity range, sensitivity rank, and version. At population size n = 18, mechanisms that require large populations are excluded rather than tuned until they produce activity.

Operational continuity is measured before the valley

Experiment R0 uses four fictional residents in a controlled decision harness. The experimental unit is one resident-history-condition-model replicate. The same underlying history is evaluated under six state representations.

Primary outcomes

No scalar is called identity. Six outcomes are reported separately.

Controls

• Blocked paired design across resident, history, task family, condition, and model.
• Held-out task variants and acceptable-action sets frozen before confirmatory outputs.
• Names, catchphrases, and formatting removed from distinctiveness evaluation.
• Prompt length matched or modeled as a covariate.
• Refusals, repairs, timeouts, invalid actions, and exclusions remain in the denominator.
• Stage A estimates measurement reliability. Stage B runs once under a frozen manifest.

R0 can support a bounded claim about behavioral continuity on registered tasks. It cannot establish subjective experience or numerical identity.

Claims come from cohorts, controls, and ablations

One persistent world is useful for operations and longitudinal observation. It is not a statistical sample. Confirmatory claims require registered cohorts with declared conditions, seeds, exclusions, metrics, uncertainty methods, and stopping rules.

Minimum analysis design

• Report famine, fission, conflict, and extinction as differences from a null demographic model, with uncertainty.
• Remove seasons, scarcity, memory, institutions, migration, disease, or LLM cognition one at a time.
• Sample global parameter space with Latin hypercube or Sobol methods, plus local perturbations.
• Define labels such as faction, war, and governance form before the runs that use them.
• Report model refusal and invalid-action rates beside behavioral outcomes.
• Publish negative, null, and unstable results under the same artifact requirements.

Launch acceptance conditions

A five-level claims ladder

Evidence at one level does not imply the next level. Public language must name the level supported by the experiment.

Phase 1 implementation status

Development starts with properties that cannot be added after histories exist. The current walking skeleton uses grain moving between containers to test canonical encoding, conservation, event integrity, named randomness, and reconstruction before ecology or cognition is introduced.

Gate 0 audit decisions remain open while Phase 1 foundation work proceeds. The next technical milestone is a headless world with weather, water, crops, physiology, and demography. It will run in large no-LLM ensembles before resident cognition is connected.