Evidence Review

---
name: market-mirror-evidence-review
description: Validate actual marketing exports and assess whether simulated predictions beat a simple baseline.
---

# Evidence review

Call `actuals_import` for CSV/JSON ads, email, website, CRM or survey exports. Map source columns explicitly, give a date and market where missing, and resolve every validation error before comparing outcomes. Use `backtest` or `holdout_backtest` with independent real experiments: training observations set calibration, held-out observations evaluate it against a no-persona training-mean baseline. State the split and sample size. An uncalibrated output is directional; if held-out persona error does not beat the baseline, keep the conclusion exploratory. Only report a probability or interval with its real calibration basis. With 10+ real survey answers, `calibrate` adds a PPI++ interval for the audience rate (valid whatever the panel's quality, assuming the real answers are a random sample of that audience), the panel's worth in real answers, and how many more real answers a target precision needs. Quote that interval and its assumption, not the raw simulated rate.

The evidence grade (A-D, "How sure is this" on the dashboard) drops one step when at least half of the simulated people are in markets the models represent less well (outside the US, UK, Canada, Australia, New Zealand and Ireland) and no real answers checked the run; quote its market note. To get real answers, give the owner the run's `real_check.md` and `real_check.csv`: the questions that decided the verdict, who to ask and how many (a +-10 point margin, or the earlier calibration plan). The sheet's `segment`, `country` and `y` columns are what `calibrate --real` reads.