arbitron 0.2.0

Multi-agent consensus rankings through pairwise comparisons

Arbitron ⚖️

par /pär/ n. Spanish — a set of two vís /viːs/ adj. Icelandic — wise, knowledgeable

Arbitron is a multi-agent consensus ranking system that determines winners through pairwise comparisons. Instead of relying on individual ratings or single perspectives, multiple agents—each with unique value systems—evaluate items head-to-head to produce robust, fair rankings.

Why pairwise? It's easier to compare two items than to assign absolute scores.

Why multi-agent? Different perspectives lead to more balanced, less biased outcomes.

Why consensus? Aggregated judgments are more reliable than individual ones.

✨ Features

• 🎯 Arbitrary Sets — text, code, products, ideas
• 🤖 Customizable Agents — unique personas, tools, providers
• 🗳️ Consistent Aggregation — diverse perspectives into rankings
• 🛡️ Bias Reduction — ensemble decision-making
• 🧩 Remixable — join data with human labels or other heuristics

🚀 Quickstart

Running your first Arbitron contest is easy!

pip install arbitron

Setup your favorite LLM provider's API keys in the environment and then run the following code.

import arbitron

movies = [
    arbitron.Item(id="arrival"),
    arbitron.Item(id="blade_runner"),
    arbitron.Item(id="interstellar"),
    arbitron.Item(id="inception"),
    arbitron.Item(id="the_dark_knight"),
    arbitron.Item(id="dune"),
    arbitron.Item(id="the_matrix"),
    arbitron.Item(id="2001_space_odyssey"),
    arbitron.Item(id="the_fifth_element"),
    arbitron.Item(id="the_martian"),
]

agents = [
    arbitron.Agent(
        id="SciFi Purist",
        prompt="Compare based on scientific accuracy and hard sci-fi concepts.",
    ),
    arbitron.Agent(
        id="Nolan Fan",
        prompt="Compare based on complex narratives and emotional depth.",
    ),
    arbitron.Agent(
        id="Critics Choice",
        prompt="Compare based on artistic merit and cinematic excellence.",
    ),
]

description = "Rank the movies based on their soundtrack quality."

comparisons = arbitron.run(description, agents, movies)
ranking = arbitron.rank(comparisons)

YAML Based Configuration

Arbitron can read the configuration from YAML files for easier management and scalability.

description: "Rank the movies based on their soundtrack quality."
comparisons_per_agent: 20
agents:
  - id: "SciFi Purist"
    prompt: "Compare based on scientific accuracy and hard sci-fi concepts."
    model: "google-gla:gemini-2.5-flash-lite"
  - id: "Nolan Fan"
    prompt: "Compare based on complex narratives and emotional depth."
    model: "openai:gpt-4.1-nano"
  - id: "Critics Choice"
    prompt: "Compare based on artistic merit and cinematic excellence."
    model: "groq:moonshotai/kimi-k2-instruct"
items:
  - id: "arrival"
  - id: "blade_runner"
  - id: "interstellar"
  - id: "inception"
  - id: "the_dark_knight"
  - id: "dune"
  - id: "the_matrix"
  - id: "2001_space_odyssey"
  - id: "the_fifth_element"
  - id: "the_martian"

You can then run the pairwise comparisons using the following command:

arbitron run competition.yaml --output duels.csv

This will create a file duels.csv with the results of the pairwise comparisons.

arbitron rank duels.csv

❓ Why "Arbitron"?

The name comes from the Old Icelandic word "val" meaning choice or selection. In Norse culture, having many choices among stories was a sign of abundance. Arbitron embodies this spirit—providing rich choice through diverse agent perspectives.

Also inspired by Philip K. Dick's novel VALIS (Vast Active Living Intelligence System), reflecting the distributed, intelligent nature of multi-agent consensus.

🏛️ License

MIT License - see LICENSE file for details.

🙌 Acknowledgments

• DeepGov and their use of AI for Democratic Capital Allocation and Governance.
• Daniel Kronovet for his many writings on the power of pairwise comparisons.

---

Margur veit það sem einn veit ekki. Many know what one does not know.