What we won’t claim
The limits, before you ask
A confident number nobody has checked against reality is the riskier purchase, not the safer one.
How do you know you're right
Today you don't, and there's no pretending otherwise. There is no published accuracy record.
What there is: reasoning you can inspect, an explicit range instead of false precision, and a stated limit on every claim. Every read says how far you should trust it, and what would have to be true for it to say more.
- LowOne witness, or witnesses that share a source
- ModerateTwo witnesses that cannot have copied each other
- ContestedIndependent witnesses disagree. Not a rung. A finding in its own right.
- HighReserved for an experiment. A geo holdout or a conversion lift test is the only evidence that can prove a lift rather than corroborate one, and it is the only thing that lifts a read above Moderate. That is a test you run, and it is the one thing this cannot work out from the data you already have.
The fair counter-question runs the other way. A confident number nobody has ever checked against reality is the riskier purchase, not the safer one.
Three things this will not say yet
Will not claim: An accuracy percentage
No confidence tier here has been graded against enough real outcomes to earn one, and a borrowed number is worse than none.
What would unlock itEnough graded months to measure a hit rate, published whether or not it flatters us.
Will not claim: That the ranges are calibrated
They are derived honestly from the evidence. Whether they contain the truth as often as they should is something to observe, not assert.
What would unlock itFrozen estimates scored against what actually happened, walk-forward, never refitted.
Will not claim: An unqualified "causal"
Without an ingested experiment a verdict is associational, and the confidence it is allowed to earn is capped at that ceiling.
What would unlock itA holdout or geo test read into the month, which lifts the ceiling for that channel only.
Each one names the condition that would unlock it. Come back and check whether the condition was met and the claim appeared.
Why say this out loud
The whole product asks you to judge the reasoning instead of trusting a number. That only works if the places the reasoning is still thin are marked on the page rather than buried.
When there's more to show, more gets claimed, and you'll see when that happened.
See for yourself
The honest version of "trust us" is "read one". Limits and all.
Read one before you decide anything.
Every channel, month by month: what each one claimed, and what the evidence actually supports.