Design · CertifiedReadout · In buildTest what works
Incrementality
“Did it actually make a difference?”
Did the change cause the difference? Sized and sealed before you spend.
01What you receive
What the change moved, measured against control
Lift readout · versus controlWhat the change moved, measured against control
The measured difference between test and control, at the resolution declared before launch — with the uncertainty around it stated, never hidden. The readout reads only the design sealed when you freeze it, so the result you see answers the test you approved. And when the honest answer is no detectable difference at that resolution, that null arrives as a certified result — same fields, same rigour, with the date it stops being true stamped on it.
02The question it answers
Bring it in plain language.
- “Did the campaign actually cause the lift — or was it already coming?
- “Is the difference big enough to be real, or inside the noise?
- “Can our data even resolve an effect this size — before we spend?
A population that can be split into test and control — regions, stores, accounts, or customers — and outcome history at that grain. Enough history to measure the noise the test must clear; checked before launch, and refused if it can't. No modelling work on your side.
03How it earns trust
Designed before it runs, checked when it reads out.
Design · CertifiedReadout · In buildDesign
Design · CertifiedThe design is sized before anything is spent. The client declares the difference that would change their decision; the platform derives, from their own pre-period variation, the smallest difference the design could actually separate; and it refuses the design if the second is larger than the first.
Refused before it costs you
A design your data cannot resolve is refused before a dollar moves — and a refusal is never billed. The readout then checks the precision again against what actually happened, so a test that fell short of the precision it promised says so, beside its result.
Sealed when you freeze it
The stores, the split, the dates and the measure are sealed when you freeze the design.
Readout
Readout · In buildSignificance from the assignment
The effect is estimated with a covariate-adjusted contrast and tested against a randomisation null built from your own data, so its significance rests on how the test was assigned rather than on a distributional assumption.
Rigour you can inspect is worth more than rigour you're asked to trust.