Metrics & analysis

The analyzer outputs lightweight metrics that help you compare configurations:

• length (chars, words)
• distinct-2 (bigram diversity)
• repetition score (repeated bigrams)
• JSON validity (when using structured outputs)

For executive decisions, you’ll often combine these with:

• human rating (1–5)
• task success rate (e.g., correct label)
• groundedness checks (for RAG)

The scripts are intentionally simple so you can modify them.