Calling a model an oracle lets a team stop checking. Calling it a tool returns the work to engineering: what goes in, how the output is verified, and who owns a failure.
A candidate is not a conclusion
Drafts, summaries, and code completions are candidates. They can be fast and they can be good. They should not skip the review that already existed.
Capability changes. Responsibility does not
A better next model does not erase the need to record this model’s errors. Turning one successful demo into an architecture is where most rework begins.
The stronger the tool, the more you need a place that can say no.
Discussion