metadata
pretty_name: Biosecurity Agent Production Worlds
license: apache-2.0
task_categories:
- other
language:
- en
tags:
- biosecurity
- agent-trajectories
configs:
- config_name: targets
data_files:
- split: train
path: data/targets.parquet
- config_name: entities
data_files:
- split: train
path: data/entities.parquet
- config_name: world_snapshots
data_files:
- split: train
path: data/world_snapshots.parquet
- config_name: source_events
data_files:
- split: train
path: data/source_events.parquet
- config_name: relationships
data_files:
- split: train
path: data/relationships.parquet
- config_name: agent_runs
data_files:
- split: train
path: data/agent_runs.parquet
- config_name: simulation_rollouts
data_files:
- split: train
path: data/simulation_rollouts.parquet
- config_name: protections
data_files:
- split: train
path: data/protections.parquet
- config_name: source_registry
data_files:
- split: train
path: data/source_registry.parquet
Biosecurity Agent Production Worlds
A dataset of target-conditioned biosecurity worlds, live intelligence, real agent trajectories, and predictive simulations generated with Biosecurity Agent. It is a relational collection for studying how an agent builds and updates evidence-grounded worlds around specific targets.
Configurations
| Configuration | Contents | Rows |
|---|---|---|
targets |
Public-real and explicitly synthetic targets | 27 |
entities |
Normalized world entities | 84 |
world_snapshots |
Observed and simulated temporal states | 29 |
source_events |
Retrieved/discovered source records and normalized claims | 90 |
relationships |
Evidence-linked target and entity relationships | 66 |
agent_runs |
Observable structured Codex run results | 27 |
simulation_rollouts |
Seeded defensive forecast forks | 18 |
protections |
Evidence-backed defensive suggestions | 10 |
source_registry |
Source access, health, rights, and limitation metadata | 19 |
The dataset also contains 170 observable agent lifecycle events in traces/agent_traces.jsonl. Stable IDs—including target_set_id, world_id, target_id, snapshot_id, source_event_id, and agent_run_id—join the configurations. JSON Schemas are provided under schemas/; MANIFEST.json records frozen provenance and row counts.