Documentation

Proving the lift

Receipts, the weekly settled verdict, and why we will say nothing rather than manufacture a win.

The question you get asked is whether it worked. This is the machinery that answers it without flattering you.

Receipts

A receipt opens the moment something ships, waits out that engine's propagation window, then classifies. Confirmed is awarded only when the before and after ranges separate. A receipt with no baseline refuses to classify at all rather than manufacture a win, and regressions render with the same prominence as lifts, including on the report you send your client.

The weekly verdict

Movement settles once a week and then stops changing, so two people opening the same client link cannot read two different stories. That is a statistics decision, not a UI preference: re-reading a dashboard daily produces a false "real change" roughly 15.5 percent of the time, against 2.08 percent when the comparison is made once per week.

Canary phrases

Plant a distinctive line on a page, and when that exact wording turns up inside an AI answer you know which page fed it. On save we scan every stored answer first and tell you whether the phrase already appears, so the plant is clean.