Evaluation

A fluent answer isn't the same as a correct one. Evaluation is how we tell the difference, and how we decide what to build next.

How we approach it

• Real-world data. We test on real code, real documents and real conversations rather than toy benchmarks.
• Grounding checks. An answer only counts if the sources it cites actually support it.
• Every change. Retrieval and answering changes are measured before they ship, not after.
• Speed is a metric. For interactive products, how fast a response arrives matters as much as what it says.

Why it matters

Evaluation closes the loop between research and product. What we measure in the lab decides what ships, and what we learn from people using our products decides what we measure next.