A world in five views
Five language-model residents share a clock, four places and a finite stock of goods. Mara can invite Nessa to take a break. Will Nessa get a turn to answer, and will the two spend that time together? We built Village Zero to study how choices made from different points of view can become coordinated action. [1] [4]
Village Zero / Networked Software Systems Laboratory / Technion
Field atlas
The village on day 24
R06 · Tick 48
The project
What would count as a society?
One resident’s decision can change another resident’s options. We want to know when the coordination that follows can survive a missed turn, a separation or a change of circumstances. [1]
Consider three ways to put an innkeeper in software. In a chat, we can ask her to describe a lively evening. In a scripted game, a designer can arrange for guests to arrive when the player opens the door. In Village Zero, an innkeeper can invite someone who is actually present. That resident has a private history, other things to attend to, and the option to decline. The designer specifies the ways an invitation can work. Whether the two spend time together is an experimental outcome. [1] [3] [4]
We use autonomy in a specific sense. The experimenter sets the geography, character profiles, available actions and clock. Once an autonomous run begins, the residents choose how to use the opportunities they get. A village that talks all afternoon and gets nothing else done is a possible result. The Long-Life protocol explicitly kept such quiet stretches in the study. [1] [7]
This question belongs to a broader line of research. Park and colleagues followed social coordination in a town of 25 generative agents. In Village Zero we can take a proposed explanation of one interaction and work backwards. What was each participant told? Was a response possible? Which change did the engine commit? Five residents are enough to study that sequence. Evidence about larger societies will require larger experiments. [1] [3]
Place does real work here. Mara keeps the tavern, Ilan trades, Talia gardens, Toren is the smith and Nessa is the village gossip. At the checkpoint above, Mara and Nessa shared the tavern. Talia was alone at the houses. To reach the tavern from there, she would need two ticks to reach the square and one more to reach the tavern. While she travels, another resident can act and an invitation can expire. Co-presence makes a local exchange possible. It leaves the choice to the people involved. [2] [3]
Architecture
Inside one resident’s turn
A language model is good at producing several plausible things an innkeeper might do. A village also needs an unambiguous answer to where she is, how many crates she has and whether her guest is still there. We give those jobs to different parts of the system. [1] [3]
- World EngineThe current locations, clock and events determine what can happen. Arrivals are applied when due.
- SchedulerA resident with eligible work gets a decision slot.
- Context BuilderLocal perception, allowed actions and selected private records become that resident’s prompt.
- Language modelThe resident proposes a structured choice and can supply words to say.
- Validation and commitThe engine checks the choice, records its outcome and delivers any resulting observations.
After a choice is committed, the world delivers observations to eligible recipients. An observation can give another resident work to attend to. If a slot remains in the tick, the Scheduler can choose that resident next. [3]
One world has to settle disagreements
The World Engine keeps locations, journeys, stocks, money and event records in SQLite. Talia can pack a harvest at home when reserves and storage permit it. A purchase requires a live offer and an eligible buyer. A journey can finish as the clock advances without a fresh model decision. Deterministic rules let us trace where an object went or when two people met. They also define the limits of the society we can observe. Residents may describe activities for which we have never implemented a state change. [2] [3]
One local model, qwen3.8:27b, supplied all five residents’ decisions in the Long-Life Run. Each turn authorizes choices for one resident. If Mara wants Nessa to do something, she has to reach her through the world and leave the answer to Nessa. Their roles, records, locations and available choices distinguish their turns. Sharing a model makes it feasible to follow the village for many days. It also gives the residents a common source of habits and errors. [1] [3] [7]
Who could know this, and when?
A resident’s history distinguishes an observation, a message from somebody else and a personal interpretation. A notebook entry or a plan can persist across turns. A recipient can remember what Mara said even if Mara was mistaken. Keeping the owner and source of each record lets us ask a precise question about a later claim. Did the resident observe the event, hear an account of it, or have only an earlier inference to go on? [1] [2] [3]
Searching those records combines word matching with similarity between embeddings, numerical representations of text. The Context Builder gives the current task and required evidence room first, then fits selected history into the prompt. The Long-Life model had a 16,384-token context window, but its native input budget was 4,096 tokens. A bounded view makes long runs practical and their inputs inspectable. The cost is scientific as well as technical. A relevant memory can exist, miss retrieval, or be retrieved and still fail to influence a choice. Each possibility calls for a different test. [1] [3] [5]
Attention takes village time
The Scheduler ranks residents with eligible work using the saved queue, waiting time and priority rules. A deadline can pass while somebody else is acting. When the schedule allows several turns in the same tick, the engine commits each result and refreshes the queue before choosing the next resident. This makes a prompt answer possible and makes scheduling policy part of the experiment. Counting refusals, for example, requires knowing how often a resident actually received a choice. [3] [7] [8]
A compact model of one turn
pᵢ = PackB(local view, legal menu, selected memories)
aᵢ ∼ Model(pᵢ)
S′ = Apply(S, i, aᵢ)
Here i is the scheduled resident, B is the input budget and S is the committed world before the choice. The model samples a proposal from the prompt. Application can record a rejection. S′ is the state after that decision, which may still be in the same tick. This notation summarizes the implemented sequence. [3]
Method
What would convince us?
First decide which record could answer the question. A delivered message establishes who said what to whom. A saved intention establishes that someone recorded a plan. To count a shared activity, we require a committed event that names the participants. [1] [4]
In R06/A a free-time event put Nessa in the queue at tick 51. One tick later Mara invited her to rest at the tavern. The invitation added a legal choice and created no new job. At tick 55 the scheduler selected Nessa’s earlier job. Her prompt contained both joining and declining. She chose to join. The resulting event names both residents, identifies the tavern and points back to Mara’s invitation. We can count one shared rest. Their feelings about each other remain an open question. [4]
Next choose what to hold fixed. A controlled snapshot can put two residents in a position to cooperate and ask whether they ever complete the task. A longer run can ask how often that position arises on its own. For a paired comparison we copy an entire world, including memories and pending work, before changing a rule. The pair is the comparison unit. Successive decisions inside one village share a history. [6] [7]
Finally, preserve the turn. Before dispatch the world database stores the full logical request. A separate transport journal records each attempt and its request hash, then the reply or failure when available. The world database records the applied result. Checkpoints identify both databases by their hashes. Recovery can use a verified saved response without drawing another one. A fresh continuation from the same point serves a different purpose. It samples another possible future. That distinction lets us both audit what happened and investigate what might have happened. [1] [3] [7] [8]
Evidence
What the village has taught us
Access to a fact and use of it are measurable separately
An early telephone-game experiment used an innkeeper, a gossip and a merchant, all played by Gemma 4 12B. A sealed blue parcel contained six silver coins. Seven details of the incident were defined in advance. Across six seeds, the merchant received either the full account, a retelling passed from Mara through Nessa, or reports from two residents who started with complementary information. The archived scoring distinguishes facts fully present in his incoming material from those fully stated in his answer. [5]
How much of the parcel story arrived?

Read the counts
| Source at the start | Incoming | Spoken |
|---|---|---|
| Full account | 42 | 6 |
| Mara → Nessa → merchant | 11 | 3 |
| Two partial accounts | 12 | 2 |
Even with the complete account, the merchant often spoke in general terms. In one direct run he mentioned the well and the delivery to Mara while omitting the six coins. His words establish what he reported. Whether he had forgotten anything would take a separate test. A single unblinded Codex reviewer scored the seven related details, and the labels have no independent validation. [5]
Two participants can contribute to a real outcome
A tavern experiment used those same three resident roles with two model families, Gemma 4 12B and Qwen 3.5 9B. It defined a three-part task. A supplier had to prepare stock, the innkeeper had to collect it, and an evening service had to consume it. In twelve controlled worlds Mara began with a prepared request. In four autonomous worlds no resident action was scripted. The table follows the original runs to their own endpoints. [6]
Where the evening preparations stopped
| Starting condition | Prepared | Collected | Served |
|---|---|---|---|
| Prepared request12 worlds | 12 / 12 | 12 / 12 | 7 / 12 |
| Autonomous start4 worlds | 4 / 4 | 2 / 4 | 0 / 4 |
The controlled runs establish that two residents’ choices can produce a shared result. The autonomous runs show how a chain can stop after useful work has already happened. In one, the records connect preparation to collection and end without a service receipt. The controlled worlds allowed pickup at the tavern. In the autonomous worlds the pickup was at the market, so serving also required a return journey. Comparing the completion rates cannot isolate the effect of Mara’s request. [6]
More opportunities change some outcomes much more than others
The Long-Life Run began with eight histories grown separately from one template. At day 24 we copied each into two continuations and ran both for 72 more simulated days, at two ticks per day. A allowed one resident decision per tick. B allowed up to three different residents to act in sequence. B delivered more messages in every pair. The median paired change in shared rests was zero. [7]
What changed with a wider schedule?

Read the counts for both schedules
| History | Messages A | Messages B | Rests A | Rests B |
|---|---|---|---|---|
| R01 | 30 | 53 | 0 | 12 |
| R02 | 9 | 22 | 12 | 13 |
| R03 | 45 | 156 | 1 | 5 |
| R04 | 67 | 301 | 0 | 0 |
| R05 | 69 | 287 | 7 | 2 |
| R06 | 7 | 56 | 9 | 1 |
| R07 | 19 | 67 | 2 | 2 |
| R08 | 31 | 115 | 1 | 1 |
B used 2,943 model calls against A’s 1,152. The recorded GPU time for processing inputs and generating outputs was about 388 versus 149 minutes. Equal village time gave B substantially more computation. The extra conversation is a measured outcome. It leaves open whether residents became more cooperative when given comparable opportunities. [7]
For Experiment 72 we returned to the moment after Ilan offered a crate and asked whether Toren would buy it if his turn came one tick later. [8]
Research direction
What would count as progress?
Village Zero can already preserve different points of view, run them forward together and reconstruct the consequences. The harder question is whether the residents can sustain useful patterns of activity as circumstances change. Four tests would move us closer to an answer. [1] [3] [7]
- A relationship that survives separation. Give a resident comparable chances to visit two people after time apart. Prior contact should predict whom they seek out even after accounting for distance, availability and who was offered a turn.
- Cooperation that can renew itself. After a successful evening, change who holds supplies or when a partner is available. Measure whether residents can initiate a new division of work and complete it over several cycles.
- Behaviour that changes with experience. From the same checkpoint, vary which of a resident’s own earlier records reaches the prompt. Keep menu, text budget and generation seeds comparable. Measure the choice and its later consequences alongside any citation to memory.
- Autonomy at a sustainable cost. Compare scheduling and retrieval policies at fixed simulated time and again at matched inference budgets. Track quiet periods, recovery, missed opportunities and completed joint work per unit of computation.
A long run can tell us that two residents met again. A well-chosen replay can ask whether an earlier meeting made the difference. Building a village where both questions have an observable answer is the reason for this project.