Research
We publish what we learn
Our research covers evaluation methodology, dataset design, and the gap between benchmark performance and real-world capability.
How we protect a company's history when training AI
When a company licenses its history to train AI, the first question is about trust, not price. Here's exactly what happens to the data, and why we treat anonymization as a matter of detail.
5 min read
Why synthetic benchmarks are holding AI back
The AI industry over-indexes on synthetic benchmarks that don't reflect real-world complexity. Models that ace leaderboards routinely fail in production.
10 min read
The data problem in agentic AI
AI agents need fundamentally different data than chatbots: multi-step interaction trajectories, tool-use demonstrations, and multi-modal workflow data. This data barely exists.
11 min read
From sandbox to production: the gap nobody talks about
80% of AI projects fail in production. The problem is not the models. It's the gap between evaluation conditions and deployment conditions.
12 min read