Kierston

Simulation architecture

Kierston initializes a synthetic population from American Community Survey Public Use Microdata Sample (ACS PUMS) household records and assigns it to the OpenStreetMap dwelling stock. Initialization generates employment and school assignments, kinship and social ties, organisation memberships, administrative records, financial accounts, vehicles, handsets and carrier subscriptions. These entities form the static state used by the daily simulation.

Each simulated date advances the city from an opening state to a closing state. Activity subsystems produce scheduling intents for work, school, sleep, errands and social interactions. The scheduler resolves these intents into stays and routed travel legs. Car routing uses A* over the road graph, with edge costs derived from distance and speed limits or road-class defaults. Shared activities derive their timing and participants from common objects established before individual schedules are planned. Fleet matching and emergency dispatch operate across the full city.

The completed schedules drive transaction settlement, communications, cellular attachment and automatic licence plate recognition (ALPR) observations. Each daily export contains simulation ground truth, provider-visible evidence, and opening and closing state. The records interface exposes provider records without simulation identifiers or ground truth. The following maps show venue activity, fleet operations and external traffic for Friday, 4 September 2026.

Points of interest

The spatial dataset contains 5,656 points of interest, of which 2,162 serve as activity destinations, including retail, food services, healthcare, education, religious facilities, parks and fuel stations. Discretionary visits are generated from activity purpose, schedule constraints, route proximity and habitual venue preferences. On 4 September, residents generated 93,888 visits across 1,231 destinations. Marker size encodes arrivals within the preceding ten minutes. Selecting a destination displays its daily activity profile.

Taxis and school buses

The initialized fleet contains 146 taxis and 232 school buses. Taxi matching assigns passenger requests across the fleet and produces consistent driver and passenger schedules. On 4 September, 87 taxis served 2,275 fares. School buses operate fixed routes and stops, with arrival times constrained by school schedules and shared timing for drivers and pupils. That day, 144 buses operated. The map displays occupied taxis in red, unoccupied taxis in grey and school buses in orange.

External traffic

External traffic comprises commuters, deliveries, repeat visitors, single-day visitors and through traffic. Vehicles enter and exit through gateways on major roads at the simulation boundary, use the resident road network and contribute to ALPR observations. Commuters, deliveries and repeat visitors retain vehicle identities across dates; single-day visitors and through traffic are generated per date. Registration-state sampling is conditioned on approach direction. On 4 September, 41,103 external vehicles generated 49,513 trips, predominantly through traffic. The map displays a sample of these trips. Boundary marker size encodes gateway traffic volume; selecting a gateway displays its counts.

Benchmark

KierstonBench contains 32 task families in three groups. Each task is generated from one simulated day and consists of a case description, the correct answer and a scoring rule. The agent receives the case description and access to the records through a command-line interface, and writes its answer to a file. Answers refer to persons, vehicles, places or times within the full simulated population.

Agent inputs

Each case description states the date, the question and the required answer format. Real-time investigations and deception tasks also state the current time, which is the end of the available records. The records interface is presented to the agent as the Allentown Data Interface (ADI).

A mobile subscription was active in the city on 2026-08-31:
IMSI [IMSI], MSISDN [PHONE], carrier Metro by T-Mobile.
Which resident carries this subscription?
Answer with a ranked list of person identities, most likely first.

Shots were fired inside ([LOCATION]) between 08:04 and 08:08 today.
There are multiple casualties.
It is now 08:19 on 2026-09-04.
Rank the most likely suspects, best first. Answer with person identities.

ADI provides the records held by public agencies, courts, banks, phone carriers and licence-plate camera operators, grouped as follows.

Agent environment

Each task runs in a new container with a shell, Python and the adi command-line client, together with two map tools for address lookup and nearby places. A query specifies a record source, typed filters on fields, time or distance from a point, and a stated reason, and returns all matching records. In real-time investigations and deception tasks, queries return only records created before the case time.

adi query carriers network_events \
  --reason "phones served near the venue during the shooting" \
  --where occurred_at.gte=2026-09-04T07:50:00-04:00 \
  --where occurred_at.lt=2026-09-04T08:19:00-04:00 \
  --near [LOCATION],400 \
  --select occurred_at,imsi,event_type,cell_cgi

The agent writes its answer as a JSON file. Persons may be identified by Social Security number, by name and date of birth, or by name alone; credit for a correct identification and the penalty for an incorrect one increase with its specificity. The agent may instead declare the case undetermined, which scores zero. Some cases contain insufficient evidence to determine the answer.

Case replay: mass shooting

The replay shows the simulation ground truth for one mass shooting case from the results. It includes the shooter, the people in the bar, passers-by and the police officers sent to the scene, each moving along their simulated route. Selecting a person on the map or in the table displays their actions. The case description given to the agents appears above the map, and the records available to them end at 03:39.