autoreason

What problem does it solve?

Autoreason provides a systematic, blind evaluation loop for subjective work by using a three-candidate tournament to compare incumbent, adversarial revision, and synthesis, judged by a panel with a Borda count. This approach helps mitigate anchoring biases and provides auditable decisions across iterations.

Core Features & Use Cases

• Iterative tournament with three agents: Critique, Adversarial Author B, Synthesizer AB, and blind judges.
• Maintains state in autoreason-state.json and logs results in autoreason-results.tsv for traceability.
• Ensures no scope creep; judges evaluate strictly on task prompt and quality criteria.

Quick Start

Initiate the autoreason loop on a target artifact to start the three-candidate tournament and record rounds.