Project Watchtower

Watchtower is a planned open platform for AI verification. The goal is to provide independent, reproducible benchmarks for AI system behavior, focusing on safety and non-determinism.

The Problem

Current AI evaluation is often opaque, relying on self-reported metrics from model developers. There is a lack of independent, standardized testing for critical failure modes in real-world deployment.

Intent

We aim to build a suite of public tests that anyone can run. Watchtower will prioritize transparency in methodology and data availability.