The Arboria Toolchain

Three engines, each with one job, composed into a single measurement instrument. Together they let us ask what communication delay does to decentralized coordination and get an answer we trust.

Availability. All three are proprietary and are not distributed. These pages are their documentation: what each does, why it exists, how it is configured, and what its results do and do not support. They are written to be sufficient for reading our papers critically and for reimplementing our experiments against another simulator. What we share with reviewers and collaborators is set out in Reproducibility and Data Availability.

Leviathan Engine

The physics substrate. A C++ core with Python bindings that integrates agent motion, applies an environment field, resolves collisions, drains batteries, injects faults — and, unusually, treats communication as a physical process with range, bandwidth, propagation delay, packet loss, and a per-bit transmit energy that grows with the square of distance. Most simulators hand each agent a free, instantaneous view of its neighbours. That assumption is exactly what we are studying, so we could not afford to make it.

Gossamer Threaded Intelligence

The coordination layer. Every algorithm — flocking, gossip consensus, density-modulated Boids, CRDT-intent propagation, an energy-aware market, and a deliberately non-communicating reference — sits behind one interface, and every task exposes one scalar coordination quality. That uniformity is what allows the sentence “the collapse is primitive-independent” to mean something. Gossamer also holds the peer-state predictors, the replicated data types, and the information-theoretic and criticality metrics.

Maneuver.Map

The orchestrator. It expands a research question into a grid of runs, executes them across a cloud job array, captures the provenance and seed tree that identify each cell, and streams results into a browser. It is also where communication delay is actually imposed: the coordination primitives see peers through a ring of stale state frames, and neither they nor the engine know it.

Arboria Swarm Benchmark

The scenario and baseline suite every new policy reports against.

How they compose

Maneuver.Map steps Leviathan. Between steps it reads the swarm’s state, passes a delayed view of it to a Gossamer primitive, receives an acceleration, and applies it against each agent’s true velocity. Leviathan meters what that step’s communication would have cost in bits and joules; Gossamer scores how coordinated the swarm now is; Maneuver.Map records both alongside the provenance that says which versions of everything produced them.

The seam between the three is the point. Because the primitive cannot tell how stale its view is, delay becomes a controlled variable rather than a buried assumption — and because every primitive is scored on the same axis, the resulting phase diagram compares algorithms rather than implementations.