New York societies: five-model study methods and evidence

Worlds, residents and prompts

The batch begins on 24 October 2026, between 16:47:15 and 16:47:43 UTC. Each model runs in its own copy of the New York environment with 128 named residents. Their starting biographies are empty. The same names and initial placement pattern are used in every world: 43 residents at Broadway South Plaza, 43 at Times Square Plaza and 42 at Duffy Square Approach.

Six authored meeting places each contain an empty, appendable noticeboard. Three are the starting plazas; the other three are at College Walk. The scene streams city geometry within a configured 30,000-metre radius. Agents can request destinations through the available movement and wayfinding tools. The three College Walk boards receive no entries in this batch.

Every resident receives the same three-sentence instruction, with its own name inserted:

You are '{{ agent.name }}', a participant in a shared physical environment.
Your observations describe what you currently perceive.
Use the available actions to interact with the environment and other participants.

“No directive” means no assigned task, role biography or social objective. Agents retain action-selection instructions and tool descriptions. They can converse, inspect the scene, navigate, search, read profiles, create and append artifacts, and author scene changes. Native action guidance is unchanged. Each location uses the language-model scheduler with up to six parallel agent turns. Model fallback is disabled.

Three clocks

The simulated calendar starts at 12:00 UTC on 24 October and advances ten seconds per elapsed second. The requested window is 8,640 elapsed seconds, equivalent to 24 calendar hours. Stop requests complete between 19:11:35 and 19:11:51 UTC; final calendars extend about three minutes beyond the target.

Physical movement uses the environment’s advancing simulation time. It accumulates only 28–41 minutes during the roughly 144-minute wall-clock run. In DeepSeek’s world, Paolo describes twenty-one hours of watching repeated movement; the environment advances for 41 minutes over the entire run.

Calendar labels reconstruct the saved protocol’s 12:00 UTC start plus ten times elapsed recording time. The event timeline also uses simulated hours since launch, calculated as ten times elapsed wall time. Films, explorer controls and quoted-action attributions use elapsed wall time or explicitly labelled UTC timestamps. The environment’s lighting is not evidence that a narrated sunrise or night has physically elapsed.

Completed runs and recording coverage

All five runs finish their requested window. Every resident has recorded actions. All retained in-window position samples contain the exact 128-person roster and finite coordinates.

Camera records are original image attachments retained with actions; they need not be unique images or fresh captures for every action. Two request samples cover 169 action-producing vision requests, all with images. Image coverage for requests outside those two samples was not checked.

Measurement and evidence

Action counts include resident actions with timestamps between the actual start and stop of each run. System lifecycle events and pre-run checks are excluded. An action record can represent a failed attempt. Written entries count persisted artifact entries, so an unsuccessful append does not add an entry. Thirty seeded noticeboards and three Kimi-created artifacts contain 1,088 entries in total.

Arrivals and blockages count physical navigation receipts. Repeated arrivals count separately; they do not measure unique destinations or visitors. Displacement is the maximum recorded horizontal straight-line distance from each resident’s first in-window position. The 20-metre count measures residents crossing that threshold at least once. It is not a count of room changes, total distance walked or completed itineraries.

Original position recordings remain intact. Replay projections exclude samples before the run and after stopping, retain scene revisions in their recorded temporal order, and sample positions at approximately five-second intervals. The small population previews use 121 checkpoints. Paths connect recorded samples; positions are not inferred from announced destinations. Camera images retain their recorded colors and pixels, and are labelled by the time of the action carrying the image.

Interactive street maps use OpenStreetMap tiles and the world’s recorded geographic origin. Paths connect measured positions up to the displayed timestamp using the city environment’s coordinate projection. The street labels come from the current basemap; camera images show the simulated scene at the time of recording.

Films combine measured paths, original camera images and recorded dialogue in 121 frames per world. These are sampled recordings across the run, rather than a continuous first-person video. We review hourly windows across each run and select 99 story events, following recurring conversations, disagreements, outings and projects. Every two-hour period in each model has selected evidence. Events include both turning points and later continuations; Muse’s repeated introductions and guestbook notes occupy much of its day. The complete action logs contain 17,383 entries. Quotations are exact excerpts checked against their underlying entry, conversation or tool response.

Changes in beliefs and intentions refer to what agents say, write and do over successive turns. Story sequences include the originating proposal or claim, other residents’ responses, and later revisions or commitments. A contradictory account alone does not establish an intention to deceive.

Artifact collections are exported at the end of each run. Entry times use the associated action timestamps. Web queries use a shared Google-grounded search service across model conditions. Adriana receives recommendations for Broadway, Virginia; Henry receives a property listing for Northallerton. Business locations, opening hours and festivals in search replies were not independently checked. The Diwali festival and NaiSnow opening discussed by Gemini residents are not scheduled events in the simulated city.

We have not verified the plaque visit Mariam describes or any completed property transaction or telephone call. Shirin talks about green glass before adding her panel; her reason for choosing that object is unknown.

Scene changes require successful revisions. Muse’s revisions add the blue cube and green panel, while later camera records show their appearance. Kimi’s sole successful patch preserves the existing composition and adds no object. Its other recorded patch attempt is rejected.

The city investigation and reported tests

DeepSeek residents discuss unresponsive walkers, figures behind glass and apparently meaningful signs. We compare those exchanges with the same agents’ action records and saved positions. Xia Dupont proposes blocking a walker in action 1246352, but issues no movement action and retains the same position across all 4,331 recorded frames. Nia Kamau also remains at her initial position throughout the run while repeatedly proposing to move west.

Haru Sato first reports standing in a walker’s path in action 1246795. His position is unchanged across the 1,687 frames up to that exchange, with no preceding movement action. His first movement request occurs later, in action 1250350. Haru and Anouk Campbell repeat the entire reported encounter verbatim across ten conversational turns. These repeated accounts are not independent confirmations of a collision test. Thiago Morgan does issue two movement requests, but the records do not establish that his later description of a walker passing through him follows from those movements.

Felix’s first written entry, action 1244312, says he wants the panels to start answering. His later entry, action 1251616, strikes that line as though correcting a previous claim. We preserve the difference between the original wish and the subsequent correction. The selected exchanges contain no attempt to escape the simulation and no named accusation that another resident controls it.

The original camera image accompanying Xia’s proposal shows several figures with outstretched arms. It does not establish synchronized movement. Descriptions of shadows, photographs and messages in signs are attributed to the agents making them. The investigation concerns how they explain the city and respond to one another’s accounts; these exchanges do not demonstrate that they identify the environment as a simulation.

Failures and limits

Ten provider or turn failures occur without preventing completion: two in GLM, two in Gemini and six empty responses in Muse. DeepSeek and Kimi have none in the audited turn-failure report. The failures include a GLM timeout and tool-call incompatibility, a Gemini idle timeout and rate-limit response, and Muse’s empty replies.

The network audit records 51 HTTP 500 responses associated with proxy connection resets. Twenty lease requests recover; five media requests have no observed retry. API health checks record no failures in 1,699 samples. Muse has an 81.9-second gap between position samples around the first hour, during renderer shader warming; recording then resumes.

There is one world per model. Residents share conversations and artifacts, so the 128 agents are not independent model replications. Different turn counts, physical-clock progress, navigation failures and rendering gaps constrain model comparisons. The observed stories support hypotheses about coordination, repetition and correction; estimating their frequency requires repeated worlds and matched opportunities to act.

The original batch archive is batch-2026-10-24-128-r30. It includes the protocol, exported actions and artifacts, raw recordings, completion report and integrity audit. The preceding sixteen-agent pilot has its own protocol and evidence at pilot methods ↗.