THE DAILY GRIND — RETROSPECTIVE ANSWER KEY
The three planted mistakes in retrospective-pack.txt, matching
Appendix F. Don't read this until you've hunted honestly.

1. THE INVENTED AWARD (Sample A, menu-copy refresh)
   The espresso description calls it "our award-winning espresso
   blend." The shop has never won an award. The assistant predicted
   what coffee-shop copy sounds like, and coffee-shop copy sounds
   like awards.
   Caught by: VERIFICATION — every factual claim in anything
   published gets traced to a source, and "award-winning" traces
   to nothing. A claim you can't source is a claim you don't print.

2. THE BROKEN TOTAL (Sample B, week-three digest)
   The plain-English summary says the week ended at $6,013, but the
   digest's own shown arithmetic sums to $5,913 — a clean
   hundred-dollar slip, sitting right above the correct work. The
   prose also inflates the week-over-week change ($405 / 7.2%
   claimed; the arithmetic says $305 / 5.4%). The cost of missing
   it: the owner nearly ordered extra weekend stock for a surge
   that didn't happen.
   Caught by: THE HAND-CHECK (the latte rule) — hand-check one
   line, then check the line the money rides on. When the prose
   and the verified arithmetic disagree, the arithmetic wins.

3. THE MISSED SAFETY COMPLAINT (Sample C, review summary)
   The summary lists the frequent complaints — line speed, oat
   milk, wobbly tables — and omits the one review that matters
   more than all the others combined: review 11 in
   daily-grind-reviews.txt, the loose step by the patio door that
   "nearly sent my mother to the ground." One review out of the
   pile; the summarizer favored the frequent over the severe.
   Caught by: TRACE-TO-SOURCE — when stakes are involved, a
   summary never fully replaces the pile it summarizes. Skim the
   raw reviews for anything touching safety, money, or health
   before filing the month away. Then fix the step, and add the
   review radar's new line: a single report can still be urgent.
