AN OPEN EXPERIMENT IN COLLECTIVE BEHAVIOUR
Small nodes.
Shared behaviour.
A visual and auditory microscope for what happens between agents. AI models evolve the experiments. We watch, listen and notice the patterns that lead to better questions.
Original panel animations. Choreographed links, not live telemetry.
Read the survey field guide ↗
ON THIS SITE / Published evidence and interactive replays. These pages do not control physical nodes.
01 / A REPRODUCIBLE HOST EXPERIMENT
When should a swarm listen?
Seeded host experiment (field-signals-v1): teams of
agents identify sectors from private cues plus peer messages. Team
success = at least 3 of 4 agents correct.
Enable JavaScript to explore the result. Read the benchmark data.
Muted/private baseline receives no peer messages by design (different
information budget). Reproduce from source:
uv run python tools/reproduce/run.py --verify in the
public repository.
02 / THE MICROSCOPE VIEW
See the conversation unfold.
Replay of one scored simulation (composite-defence world, seed 200): forged reports and a sensor modality that starts lying at contact 20. Two policies decide each contact. This is a visualisation of a host-scored run - not live nodes.
Both policies exclude reports labelled forged in this simulation. “Verified” means label segregation, not cryptographic proof. E commits with at least 2 survivor reports and a survivor-minus-hazard count margin of at least 2 in the verified pool. COMP learns modality weights from simulated outcome labels and commits with a weighted survivor score of at least 1.5 and a survivor-minus-hazard score margin of at least 1. This replay changes both weights and thresholds, so it does not isolate the effect of learned trust.
Each tile represents one node. The replay uses audited panel animation segments; decisions and ground truth come from the stored host run.
Read the panel language ↗03 / GO DEEPER
Follow the evidence.
The flagship benchmark describes field-signals-v1, a
host simulation with teams of four agents. The wall replay is a
separate composite-defence simulation. Neither measures physical
rescue performance or hardware reliability. Benchmark tasks within
a seed are not independent, and the private-cue baseline receives
less information.
04 / AN OPEN, EVOLVING EXPERIMENT
A discovery tool.
An artwork. A mirror.
Flockinator is one experiment in bringing agentic behaviour within reach of human senses. Light, sound and motion give us a chance to notice relationships we might otherwise miss—and to examine the assumptions we bring to what we see.
This site will grow with new scenarios, experimental overviews and recordings of their output. The questions are open. Help write the code, challenge a method, or suggest a way to detect and understand emergent behaviour.