---
title: "EvoNet"
url: https://manvsmidi.com/projects/evonet/
category: experiment
status: writeup
date: 2026-09-05
tags: [python, simulation, data, graphs]
related: [collusion-wiki, turfmaster, bramble, life, insert-disk-2, conclave]
---

# EvoNet

> An agent-based simulation of an arms race between self-replicating spreading agents and a defender, across three network topologies. 330 reproducible runs and a paper-style write-up.

Experiment · Write-up · 2026-09-05

- [Source](https://github.com/sadovsky/evonet)

EvoNet asks what evolves when a population of spreading agents is hunted by a defender, and how much the shape of the network matters. It's deliberately abstract: an agent is a 16-float parameter vector, not code, and detection is a probability, not a scanner. That makes topology, defender memory and perturbation rate clean independent variables, and every run is exactly reproducible from a config and a seed.

## The model

Six of the sixteen parameters do something: spread aggressiveness, stealth, self-perturbation rate, mixing rate, a preference for high-degree neighbours, and payload diversity (variance on spread success). The rest are inert and drift. The perturbation rate is itself a parameter, so how fast agents change is also under selection.

Each tick: agents try to spread to neighbours, co-located agents may mix parameters, the defender removes some, background removal takes a few more, and overfull nodes are culled. The defender is either **static** (detection falls with stealth) or **adaptive**, keeping a decaying histogram of the stealth values it has seen and punishing whatever is currently common.

The main experiment is 3 topologies (Erdős–Rényi, Barabási–Albert, Watts–Strogatz) × 2 defenders × mixing on/off, 10 replicates of 2000 ticks each.

![Line charts of defender memory mode against population mean stealth for six adaptive treatments](https://manvsmidi.com/_astro/defender-chase.51jKu9Tq.png)
*The adaptive defender's memory chasing the population's stealth.*

## What came out

- Stealth rises with detection pressure: 0.08 with no defender, 0.44 static, 0.64 adaptive. The adaptive defender keeps the population spread out instead of converging on one strategy.
- Topology changes how big a population survives, not which strategy wins (p = 0.94 for stealth). The scale-free graph sustains the smallest population, the reverse of the classic spectral-threshold ordering.
- The apparent benefit of parameter mixing was an artefact of the fitness proxy inside the mixing operator. With a neutral replacement rule it disappears.
- A multiplicative signature term leaves agents at maximum stealth permanently undetectable; an additive one closes that gap.

## Running it

```bash
pip install -r requirements.txt
python cli.py experiment configs/default.json --experiment_name main --figures
```

About ten minutes on four cores. The full write-up is in `paper/evonet_paper.md`.

## Connected projects

- [collusion-wiki](https://manvsmidi.com/projects/collusion-wiki/index.md): linked
- [TurfMaster Pro](https://manvsmidi.com/projects/turfmaster/index.md): shared tags: simulation
- [Bramble](https://manvsmidi.com/projects/bramble/index.md): shared tags: graphs
- [LIFE](https://manvsmidi.com/projects/life/index.md): shared tags: simulation
- [INSERT DISK 2](https://manvsmidi.com/projects/insert-disk-2/index.md): shared tags: simulation
- [Conclave](https://manvsmidi.com/projects/conclave/index.md): shared tags: graphs