← All posts
Build the eval harness before the product
A practical pattern for evaluating a high-stakes clinical-copilot prototype before tuning prompts.
Consider a fictional clinical-copilot brief: the instinct is to start prompting and see what comes out. A safer prototype plan starts with the evaluation harness, before the product has a working feature. This article is an illustrative engineering scenario, not a client case study.
Why eval-first
LLM systems fail silently. A model can hallucinate a diagnosis code with the same confidence it states a correct one. Before writing the retrieval pipeline, define a representative labeled set using synthetic or properly governed data and document how every example was produced.
What the harness actually checked
Three useful checks are factual grounding against the source transcript, omission of required clinical fields, and tone drift over long sessions. Each prompt or retrieval change should run against the full suite before it reaches staging.
The payoff
The goal is to expose and document failure modes before a prototype reaches clinical review. That creates a more useful conversation than “we think it is pretty good,” while still requiring domain validation, privacy review, and appropriate human oversight before real-world use.