A symptom tracker that compares every factor against every symptom will always find something. Compare ten trackers across four time lags and you have asked several hundred questions of a few months of noisy data; a handful will come back looking like answers. Earlier builds of Longnote used a single bar for surfacing links, scaled to how much data there was. Against seeded synthetic logs with known, planted effects, that bar promoted about three false links out of every five at ninety days. That is not a tracker being honest. So the rule changed.
The three standings
Observation. A fact about one tracker’s own entries: logged on 18 of the last 30 days, average 4 of 10, worst on Tuesdays. It cannot be wrong, and it claims nothing about why.
Early sign. A possible link between two things with under sixty days of overlap. It is worded as a hunch, carries its day count, and is limited to same-day or next-day timing, because later lags were mostly chance in the seeded data. Only one early sign is shown at a time; in testing, a second one was wrong more often than it was right.
Pattern. A link that has at least thirty days of overlap, survives a Benjamini–Hochberg false-discovery check across every pair and lag the app scanned, at a rate of one in ten, and still points the same way in the range on screen. The check deliberately counts every comparison, including the ones that failed, because skipping the losers would understate the family and let noise through.
A fourth standing, in lists only
Hint. Since 2 October, the browsing lists, the weather sheet and Moves with it, are looser than Highlights. They show every unconfirmed same-day or next-day link that clears the bar, at any day count, so an early sign there is not limited to one. Where one side of a comparison rests on only three or four days, the link is shown as a hint, with both counts in its caption: Hint · 3 days with, 12 without. A hint never reaches Highlights, the report, a Home tile or the weather line on a tracker’s page. Lists are for browsing your own record; Highlights are what the app is prepared to put its name to.
The evidence words
Beside each link, a word: limited data, fair evidence or good evidence. Limited means under thirty days of overlap, or a neighbouring lag that actively disagrees, or a comparison resting on fewer than eight days on one side. Good means at least sixty days, clean, and a lag-adjusted p-value small enough to survive a scan of a hundred pairs. Everything else is fair. A link in the limited tier never reads as strong, whatever its correlation, because “strong link on twelve days” is the trap this whole design exists to avoid.
What it looks like
Weather and factor cards end with the line associations, not causes. An early sign says “so far”. A comparison shows both counts: logged on 11 of 15 rainy days, and on 9 of 15 dry ones. And when there is not enough data yet, Highlights say exactly that, “No patterns yet”, because that is a result, and often the most useful one.
What it is tuned against
Synthetic logs with planted effects, so the right answer is known, and a side-by-side with another tracker’s insights over the same sixty days. The sharpest finding from that comparison was “good evidence” being awarded on pure noise, which is why two months of overlap now stops at fair when the link is weak enough that a scan this wide would turn it up by chance. None of this makes the app’s statements true. It makes them honest about how sure they are, which is what you can actually ask of a diary.