These notes show what is being tested, what is already defensible, and what still needs a sharper question.
Not every useful answer is a finished job. The current working definition is practical: a task is complete when the original decision, deliverable, or action can proceed without another round of work being necessary for its stated purpose.
A model can produce a genuinely new angle without creating a new obligation for the user. The distinction matters because novelty feels valuable even when the original task is already done.
Traditional project management assumes that scope expands through human requests. AI-assisted work adds another source: the system itself can propose attractive extensions. A good workflow needs a way to capture them without accepting them automatically.
Cars make the difference between engine cost and journey cost intuitive. They can also hide important differences: AI work is not physical transport, and human judgment is not a passenger. The analogy is a bridge, not a proof.
The useful benchmark is a fixed task run through different combinations of model, prompt, review, and stopping rule. Measure time to useful completion, not just output quality or token usage.