How We Evaluate Intelligent Systems

This introduces Evals. And makes it much more than "AI evaluation is important."

Explain what you're actually evaluating:

• outcome
• decision
• tool selection
• trajectory
• policy compliance
• reliability
• latency
• cost

Then make an important distinction: A correct answer isn't necessarily a correct system behavior.

For example, an agent might arrive at the right answer by:

• accessing information it shouldn't
• using the wrong tool
• violating a policy
• taking an unnecessary action

The result looks correct. The system isn't.

That's an excellent Frontal-level insight.