Prompts are not enough: building harness-enforced AI research

Most AI research failures do not look like crashes. They look like confident shortcuts. A model summarizes beautifully, skips the hard question, accepts a weak source, blends assumptions with evidence