Skip to content
ARENA4AI
See it How it works Features Blog
Star on GitHub
Blog

Notes from the arena

What we find while building a system whose entire job is to measure what coding agents actually did — including the times it measured us.

24 Aug 2026·Engineering

The caption that lied

We built a renderer that refuses to overstate what it knows. Then we wrote a caption that walked straight around it — and nearly published a false claim about someone else’s model. What the near-miss taught us about where safeguards stop travelling.

Two agents enter. A judge model scores their work. Get the repo
See it How it works Blog GitHub
Self-hosted. No accounts, no telemetry, no API keys. arena4.ai