← Every exhibit
The measurement, and the traps it must survive
Four steps that make the exposed-versus-untouched split mean something: an exposure tag set before close, no revenue ranking, a placebo split that must come back quiet, and enough accounts that the answer is not noise.
| Step | What it guards against | What the seeded simulation shows |
|---|---|---|
| Tag integration exposure per account, before close | Without the tag, any later comparison is folklore | Not a field in any system; an afternoon to reconstruct |
| Never rank on pre-close revenue | Regression to the mean, read as damage | The top decile of 400 accounts reads −15.1 pp with no effect present, in 73% of runs |
| Run the placebo split first | A measurement that alarms on random halves | Random halves recover the true noise floor at every spread tested (0.20, 0.45, 0.80) |
| Respect the base-size floor | Twenty-point swings that mean nothing | 95% noise range: −33.1/+34.2 pp at 100 accounts, −24.2/+24.4 at 200 |
Swipe or scroll horizontally if the table is wider than your screen.
Reference & Evidence
Source: integration-exposure-null.py: 400 accounts, 2,000 trials, seed 20260807; published beside the pair's claim ledger and reproduces bit for bit. Author's own design; the simulated numbers are the null, not a finding about any deal.
Related exhibits
-
The sign is a decision: the tested splits
From the essay The integration plan spends what the deal bought.
-
The base rate, drawn: acquirer returns drift negative
From the essay The integration plan spends what the deal bought.
-
The revenue-management control record
From the essay Revenue management versus dynamic pricing