Prompt and evidence scoping versus model scaling

A repair harness gives its language model one judgement and no authority

Maintenance work where a wrong fix returns a plausible answer needs an oracle before it needs a model, and one production deployment now reports what happened when the model was given a single judgement and nothing else.