The fetch runbook searches the last 24 hours across the AI stack each morning. For model releases, API updates, and research papers, it reads the primary source directly — the official blog post, arXiv abstract, or GitHub release page — not secondary coverage.

Technical claims (benchmark scores, pricing figures, context window sizes, parameter counts) are only included in a briefing if they appear verbatim in the primary source snippet. If a figure is not in the prep file, the briefing uses qualitative language instead.

The curate runbook does not use web search during synthesis. The prep file is the only permitted data source. This keeps every claim traceable to a specific fetch run.