An evaluation harness before an LLM feature ships
This page is a placeholder. It names no customer and reports no metric.
The situation this pattern fits
Someone wants a language-model feature in a product: a draft, a search box over private documents, or a support reply. The demo works. The open question is what happens on the inputs the demo did not show.
What the work usually includes
- A written task definition and a set of cases, including cases that should refuse.
- An evaluation that runs before a release, not only a spot check in a notebook.
- Guardrails for prompt injection and for data the feature should not reveal.
- A decision recorded for what is still too risky to ship.
What a real write-up will add
The same rule as the other placeholders: no customer name, no quote, and no performance number until a real project is approved for publication.