Ask HN: How are engineers evaluating non-deterministic ML/LLM based deployments?

[zuck_vs_musk]

So, we process data as well as documents from various sources, then,

  - convert all of its text (using different OCRs)
  - pass it to LLM models - depending on the customer, it can be a cheaper model, and we do have model fallbacks

How do engineers evaluate such systems?

  1. New models & new libraries are coming all the time
  2. Even a third-party's deployment model will change over time and might improve/regress our systems

Any good approach for writing evaluations for these?