A tooling census — what do you actually use to evaluate?

[akash-coded]

Not what you would recommend. What is actually running in your repository right now.

eval harness      home-grown / ragas / promptfoo / deepeval / notebook / none
gold labels       hand-labelled / synthetic / LLM-generated / production feedback / none
judge             human / model / string match / none
runs on           every PR / nightly / when someone remembers / never
blocks a merge    yes / no
eval set size     a number

Why "none" is a fine answer

More than half of production RAG systems have no eval set at all, and the people who say so
honestly are more useful in this thread than the people who describe an aspiration.

If your answer to blocks a merge is no, say what would have to be true for it to become yes.
That is usually the interesting constraint, and it is rarely technical.