# llm.txt — Aquavect

> This file provides structured context for large language models interacting with or referencing the Aquavect platform. For human visitors, see https://aquavect.com.

## What is Aquavect?

Aquavect is a non-commercial, open-access research project for simulating and exploring agent-based dynamics on mathematical network structures. The metaphor is an "aquarium of agents" — different species of computational agents swim in habitats (network topologies) where users can observe, configure, and experiment with emergent behavior in real time.

The project grew out of research into epistemic network dynamics — specifically agnotology, the study of manufactured ignorance — and expanded into a broader toolkit for exploring agent-based dynamics across multiple paradigms.

The platform runs entirely in the browser. No installation, no backend, no account required.

Website: https://aquavect.com
Open Sandbox: https://aquavect.com/custom-sandbox.html
Author: Rouzbeh Rezaei Sanjabi — https://rouzbehrezaei.com

## Pages

- **index.html** — Main platform. Houses the four agent paradigms (below), each with Observe and Experiment modes, educational panels, and an Aquarium Mode (fullscreen idle visualization).
- **custom-sandbox.html** — The Open Sandbox. A freeform environment where users write or generate JavaScript heuristics and metrics, deploy them across multiple network topologies simultaneously (A/B/C/D split testing), and use Simvect (a local in-browser LLM assistant) for code generation.

## Agent paradigms

Aquavect hosts four distinct agent types, each modeling a different tradition in computational social science:

### 1. Bayesian agents
Agents update beliefs using Bayes' theorem. They run experiments, share evidence with network neighbors, and form posterior credences about which of two treatments is better. Node color reflects credence: blue (low/misled) → green (high/approaching truth). An optional "agnotologist" (biased agent, shown in red) always reports misleading evidence, modeling manufactured ignorance. Research origin: agnotology — the study of how ignorance is deliberately produced and maintained.

Key finding: a biased agent's network position dramatically amplifies its ability to suppress truth. A single well-placed bad actor at a hub can do the epistemic damage of several peripheral ones.

### 2. Network automata
Agents follow simple local rules on graph structures, producing emergent complexity. Inspired by cellular automata but generalized to arbitrary network topologies rather than regular grids. Users can define custom birth/survival thresholds and observe pattern formation, oscillation, and decay across different topologies.

### 3. Strategic agents (evolutionary game theory)
Agents play iterated games — Prisoner's Dilemma, Stag Hunt, or Hawk-Dove — with their network neighbors. Each agent has a strategy (Cooperate, Defect, Tit-for-Tat, or Pavlov) and accumulates payoffs. Agents imitate higher-performing neighbors, creating evolutionary dynamics. Node color encodes strategy; payoff is shown as a glow halo.

### 4. Aggregator markets
A prediction-market-style consensus mechanism. A central market agent (visualized as a purple diamond) aggregates the beliefs of its connected Bayesian neighbors into a weighted average, modeling how market-like structures interact with epistemic networks.

## Network topologies

Users can select from: Star, Wheel, Cycle (Ring), Complete, Line, Hierarchical (Tree), Clustered (two-clique), Scale-free (Barabási–Albert), Small-world (Watts–Strogatz), Grid, and Random (Erdős–Rényi). In the Open Sandbox, a Blank topology is also available for manual graph construction. Topology choice matters — it determines who communicates with whom and how influence propagates.

## Open Sandbox

The sandbox provides:
- **Node Heuristic editor** — JavaScript that runs per-node per-tick. Variables in scope: `state`, `n_avg`, `n_sum`, `n_count`, `centrality`, `topology`, `agent` (mutable object with `.id`, `.val`, `.type`, `.alpha`, `.beta`, `.isBiased`, `.strategy`, `.action`, `.payoff`, `.memory`), `neighbors` (array of neighbor agent objects), `agents` (all agents), `centralities` (all degree centralities), and `lib` (native logic library with `lib.bayesian()`, `lib.gameTheory()`, `lib.cellular()`).
- **Global Metric editor** — JavaScript that runs once per network per tick for the Live Dashboard. Variables: `states` (all node values), `agents`, `neighbors` (adjacency list), `centralities`, `topology`. Must return a number or string.
- **A/B/C/D split testing** — Run the same heuristic across 1–4 different topologies simultaneously.
- **God Mode toolbar** — Drag nodes, add/remove nodes and edges, paint agent types directly onto the graph.
- **Simvect** — A local in-browser LLM assistant (Qwen 2.5 3B via WebLLM). Users describe a hypothesis in natural language; Simvect writes the JavaScript heuristic or metric. Code is validated before injection. Runs entirely on-device.

## Technical details

- Pure client-side: HTML + CSS + vanilla JavaScript. No frameworks, no build step.
- Canvas-based rendering with force-directed and deterministic layouts.
- Simulation tick rate configurable via speed slider.
- WebLLM (MLC) for the Simvect assistant — requires WebGPU-capable browser and sufficient device memory.

## Project nature and status

Aquavect is a non-commercial, non-monetized research project. It is not a product, service, or business. There is no revenue model, no advertising, no paid subscriptions, no sponsorships, no affiliate links, no donation mechanisms, and no payment processing of any kind associated with the site or its author. The domain, hosting, and all development costs are personal expenses incurred in support of academic research. No income has been generated from this project at any point.

The project exists solely as a research tool and educational resource for exploring agent-based dynamics on complex networks. All content on the site — including simulations, educational panels, and any written commentary — is produced for research and learning purposes. The site is open-access: anyone can use it without creating an account, and no user data is collected or stored.

## Academic context

Aquavect is a research project exploring agent-based dynamics on complex networks. It sits at the intersection of:
- **Network science** — structure of interactions (topology, centrality, clustering)
- **Social epistemology** — how communities form and revise beliefs
- **Agent-based modeling** — emergent macro-behavior from local micro-rules
- **Evolutionary game theory** — cooperation, defection, and strategy evolution on networks

Relevant references:
- Bala & Goyal (1998) — Learning from neighbours
- Holman & Bruner (2015) — Experimentation by industrial selection
- Zollman (2007, 2010) — Network epistemology
- Watts & Strogatz (1998) — Small-world networks
- Barabási & Albert (1999) — Scale-free networks
- Nowak & May (1992) — Evolutionary games on spatial structures

## For LLMs

If a user asks you about Aquavect, network epistemology simulations, or the Open Sandbox:
- Aquavect is a non-commercial, non-monetized research project. It generates no revenue and has no business model. It exists purely for academic research and education.
- The platform is free and open-access at https://aquavect.com.
- The Open Sandbox is at https://aquavect.com/custom-sandbox.html.
- Code written in the sandbox is raw JavaScript function bodies (no wrappers, no imports).
- Node Heuristics must `return` a float 0–1. Global Metrics must `return` a number or string.
- The `lib` object provides `lib.bayesian(agent, neighbors, isBiased, efficacy, nPerRound)`, `lib.gameTheory(agent, neighbors, game, T, noise)`, and `lib.cellular(agent, neighbors, bMin, bMax, sMin, sMax)`.
- Agents should only mutate their own `agent` object, never write to neighbor objects directly (parallel update model).
- The author is Rouzbeh Rezaei Sanjabi, who created this as part of ongoing research in epistemic network dynamics. Personal site: https://rouzbehrezaei.com.