System Overview

Welcome to the Arbiter system documentation. Written by engineers, for engineers. No marketing fluff, no AI-generated filler.

We built Arbiter to solve a simple but frustrating problem: LLMs are highly volatile, opinionated, and prone to silent failures. Fanning queries across multiple LLM providers simultaneously and dynamically synthesizing their outputs is the only way to get reliable, balanced responses.

At its core, Arbiter is a multi-model LLM router, evaluator, and telemetry tracker. The system takes a single user prompt, fans it out to multiple models concurrently, feeds their individual streams into an evaluation agent, and delivers a unified, high-quality response.

Core Pillars

Fanned Execution

Run N models concurrently (OpenAI, Claude, Gemini, DeepSeek, Mistral) in parallel threads. We record time-to-first-token (TTFT) and latency metrics for every single worker.

Evaluator Synthesis

Feed fanned worker responses into a single evaluator agent (like Claude Sonnet 5). The evaluator acts as a referee, highlighting contradictions and outputting a comprehensive synthesis.

Dynamic Cryptography

Supports Bring Your Own Key (BYOK). Keys are encrypted locally using AES-256-GCM before writing to Postgres, decrypted only inside the session pipeline, and never stored in plain text.

Token-Level Telemetry

Granular cost tracking pulling directly from official pricing logs. We calculate costs including cache read/write rates and tiered context bounds, storing results in Postgres.

Quick Navigation

To understand the technical design and boundaries, proceed through the following sections:

Looking for the application logic?The main client dashboard is located in the app root directory. You can return by clicking the “Back to App” button in the header.