Rubric evals
Grade open-ended output with a model, a rubric and two examples.
LAST UPDATED
When there is no single right answer, grade against a rubric.
Write the rubric
Name one failure, say what a pass and a fail look like, and give one example of each.
Pin the grader
Tracewell records the grader model and version with every score. Changing the grader creates a new baseline.