An app that is allowed to tell you nothing

Almost every habit tracker will tell you that you are doing well. It is the obvious product decision: encouragement retains people, and a cheerful number is easier to ship than an honest one.

The problem is that it makes the numbers meaningless. If the app always finds a pattern, finding a pattern tells you nothing. And if the streak counter is the only feedback, you end up optimising the counter rather than the thing it was supposed to stand for.

So Crescendo does something less comfortable. It fixes its hypotheses in advance — sleep against training, meditation against the next day's mood, reading against focus — and then tests them properly, with Bonferroni correction because testing several things at once makes a false positive likely by construction. A finding is shown only if it clears three floors at the same time: at least twenty paired days, a correlation of 0.25 or stronger, and a corrected p-value of 0.00625 or lower.

Two of the hypotheses in the demo dataset are deliberately weak, and are correctly rejected. That is how you can tell the filter is doing something rather than decorating a foregone conclusion.

What that costs

For the first three weeks the Insights screen says, in effect, not yet. Twenty paired days is a real threshold and there is no honest way around it. We could show you something anyway — a trend line through eight points, a percentage with no confidence interval — and it would look better and mean less.

About your journal

Journal text never leaves your account. Not to an advertising platform, not into a training set, and not into our own analytics — the event we record when you write is a word count, not the words. The AI reflection feature sends the specific entry you asked about and the statistics we already computed, and retains neither.

Mood and sleep are treated the same way. They are health data, and they are not forwarded to an ad platform under any consent setting, because being permitted to do something is not the same as it being reasonable.

How it is built

The statistics are tested against independently known answers — Anscombe's quartet, published t-distribution critical values — rather than against whatever the implementation happens to return. The demo dataset plants real relationships and a separate script checks that the analytics rediscover them with no knowledge of the generator. Every colour in the interface was solved for contrast rather than chosen by eye, and audited in both themes.