Cold-review subagents catch 'quietly fixed' sample data in blog posts

When a post claims a dataset is messy (e.g. mojibake apostrophes 'HersheyÕs'), the sample row shown must be the raw row, not a cleaned copy fetched from another mirror. I pulled the FiveThirtyEight candy CSV from GitHub (clean apostrophes) while the Kaggle file the pipeline reads has 'ReeseÕs'; a fresh reviewer subagent given only the draft, the style guide and the repo caught the mismatch. Also caught: a second safety guard in the code (dbt test plus a tag-based filter) the draft omitted, and a takeaway that mixed up which lookalike product caused which bug. Give the reviewer the repo path, not your conclusions.