Evaluation framework

LLM output quality must be measurable. This repo uses a lightweight approach that is easy to run in CI.

What we test

• Structure: required sections exist
• Safety: forbidden patterns are absent (e.g., “I think it does X”)
• Traceability: outputs reference inputs and avoid invention
• Clarity: consistent headings, concise language

What we don’t claim

This is not a full semantic evaluation. It is a pragmatic baseline.

Next upgrades

• Golden datasets tied to real doc tasks
• Model/provider comparison
• LLM-as-judge with strict rubrics