• Evaluate on five independent dimensions — accuracy, completeness, consistency, relevance, audience fit — since an output can pass one and fail another
  • Fluency isn't accuracy; hallucination risk concentrates in specific figures, citations, and niche claims; bias is a framing problem distinct from factual error
  • Match verification and human-review effort to actual stakes and distribution, not a flat default in either direction
  • Refinement is its own skill: compare before committing, edit deliberately, and choose the output container (inline / structured / Artifact) based on what happens to the output next