data-verificationlisted

# Data Verification Your arithmetic is probably fine. That is not where analyses go wrong. ## How to use this **Trigger:** you are about to state a number, a comparison, or a causal claim that came out of data and that someone will act on. "The strategy makes money." "Churn is worse on the enterprise tier." "Fees are what killed it." "Cohort B converts better." **Cost:** questions 1 and 2 are the routine minimum and usually take a few minutes. They target the failure modes behind ~79% of the audited corpus below, which is a statement about where those errors came from, not a measured catch rate against them. Do not skip them because the analysis felt simple; every incident in that corpus felt simple. **Do it BEFORE you write the conclusion, not after.** Answering these after you have stated a finding turns into serial public correction, where each message walks back the last, which costs more trust than one wrong answer because it makes the reader into your QA process. **Output:** finish by reporting the number in the four-part form under "Reporting a verified number", including which questions you skipped. ## The finding this is built on An audit of ~130,000 agent messages across four agents looked for every case where an agent reported a data conclusion that was later retracted. 49 verified incidents. The taxonomy: | root cause | share | | ------------------------------------------ | ----: | | population / coverage bias