The model doesn't decide
In Scout the obvious move was to let an LLM recommend the matches. I dropped it: it can't be explained, it isn't reproducible and it doesn't learn from your feedback.
A small Bayesian model running in the browser makes the call. GPT-4o does two things, neither of them deciding: it turns one sentence into preferences and writes up the justification the model already computed.
If the API goes down, the app still classifies all 72 matches. The LLM improves the experience; it is never a single point of failure.
