Module 02 Summary

What this module established

Six dimensions, six bars, six verdicts - a composite absorbs the one that is failing and needs weightings you cannot defend. Grounding has three states, and correct-but-uncited is the dangerous one: the retrieval half contributed nothing and general knowledge happened to match.

Carry forward

• Grade correctness on facts present or absent. It handles paraphrase, separates incomplete from wrong, and two people can agree on it.
• Format is the one dimension cheap enough to check on every production response, and the one most likely to break between eval runs.
• Measure correctness for each configuration before pricing it. The last point of quality is the most expensive and is sometimes negative.

Before moving on

Move on when six dimensions are measured separately and you can name the one an average would have hidden.

Mark complete