A model that survives every result is not a model.
It is a story. This one has a date on it. Every claim this app makes about somebody is sealed when it is written, opens itself on a date nobody can move, and is counted below whichever way it landed.
The self-improvement industry does not keep this page. Neither, it turns out, does the one consumer category that actually predicts: period-tracking apps get the fertile window right about 21 to 22 percent of the time and the exact day about 8 percent, and not one of them publishes that. Worsfold et al., 2021
What a claim is
Sealed when it is written. What is public is the subject, that a claim exists, and the month it opens. The content and the thing it will be measured against are not shown to the person it is about, and that is not coyness: tell somebody you expect them to go flat on Thursdays and they will watch Thursdays. A claim that changes the behaviour it predicts can never be wrong, which is the same as never being a claim.
Measured against something collected for another reason. The endpoint is one of a short, fixed list of readings this app already takes on its own schedule, for reasons that predate any claim about anybody: the daily mood tap, how a week of a commitment came out, a re-read of the wheel. Nothing starts being collected because a claim was made about it, and the list does not grow when the claim changes.
The bar is written before we look, and we cannot move it. What counts as the claim holding is a constant stored with the sealed row and identical for every person under the same template. It is copied at the moment of sealing and a row whose contents no longer match their own hash is left open forever rather than given a verdict, because writing one would launder an edit into this page.
Nothing about the verdict involves a language model. Code reads the endpoint on the date, compares it to the bar, and writes one of the words below. There is no step at which anything gets to explain a result it did not like.
And it opens silently. No notification, no tap to reveal. A sealed envelope with an alert on it is a loot box, and if opening it quietly costs us attention then that is the product working.
The four ways one ends
- heldThe claim survived the test. It stands until something else kills it.
- wrongIt did not. The claim is struck and what replaced it is on the record.
- not enough to settleThe data needed to test it was not there. Not a pass and not a failure, and counted separately so a thin week cannot be read as a win.
- still sealedWritten, dated, and not yet open. Nobody has seen it, including us.
The third one is the reason to trust the other two. A ledger containing only clean falsifications is doing the thing we accuse stories of doing, so a claim whose data never arrived is filed as exactly that and never quietly resolved in our favour.
The record
Nothing has been claimed yet, so there is nothing here to have been wrong about. The first entries arrive when the first sealed claims open.
Counted per template rather than added together, because a template that is harder to settle would otherwise drag another one’s figure around. That is hedging achieved by averaging, and it is the failure this page exists to not have.
This is not a scoreboard. It is the part of the method that can embarrass us, and it is public for that reason.