A nightly job that recommends better local models

I have a command named "makeImage" that spelled "Grünerløkka" wrong. It wrote "Grüre Lløkka". So I set up a scout that tests other models for me, and a new one got it right both times!

My PC now looks for new image models every night. A Claude task picks one, tests it against the model that I now use on six hard prompts and writes me a recommendation. A decider model, a local vision model, looks at each pair of pictures. The new model, MiniMax H3, won 3 of 6 tests. That's on one PC with two tries per test, so it's a first look.

The decider model that I now use liked the old model's coq au vin card better but I liked the new one's. So I decided, not the decider model.

The decider also favours the first picture it sees, so I ask it in both orders and count only the answers that agree.

I wrote a blog entry about how it works. You should read it if you want to build your own: https://www.digitaldias.com/blog/2026-11-14-local-image-model-scout/