Building evals

Introduces core eval components (code-based grading, human grading, model-based grading) and how they fit into an iteration loop.

• Upstream notebook: misc/building_evals.ipynb

What to focus on

• Defining metrics that match product requirements
• Separating dataset creation from scoring logic
• Making evals runnable in CI (fast/targeted)

Run locally

make test-notebooks NOTEBOOK=misc/building_evals.ipynb