Most period trackers start the same way: "When did your last period start?" Then "When did it end?" You pick two dates from a date picker and the app stores a range.
We went the other way. In Mori the calendar is the input. You tap days. That's it.
That sounds like a small UX difference. It ends up changing the data model, the correction flows, and a good share of the edge cases in the app.
What a marked day means
There is one fact per day: recorded or not. No start field, no end field, no "period in progress" state.
A period is the longest run of consecutive recorded days. Length is a derived value: count the days. The gap to the previous run is the cycle length. Nothing else is stored.
Everything a user thinks of as "editing a period" becomes an edit to a set of booleans.
Three things this buys you
Merging is free. Tap Monday through Friday and you have one period of five days. No dialog asks whether Friday belongs to the previous entry or to a new one. Adjacency decides.
Splitting is free. Say you marked six days but the last three were spotting you logged by mistake. Delete day four. The run breaks into two, and lengths and gaps are recalculated from the set. There is no "delete period" flow, because there are no period objects to delete.
Corrections look like edits. Fixing a wrong range in a date picker usually means deleting a record and re-entering it. With day-level state, fixing a mistake is the same gesture as making it.
The edge cases that forced the design
Day-level state pushes decisions to the read layer, which is where I want them:
- Two non-adjacent taps. Two runs of one day each. We keep both and lower confidence instead of quietly joining them.
- A gap far outside the user's own history. The record stands exactly as entered; the prediction interval widens and says so.
- A very short run. Same treatment: keep the fact, distrust the input.
The rule we settled on: never rewrite what the user recorded to make the statistics look cleaner. If the numbers are odd, confidence drops, not the data.
What we gave up
- No notes attached to "this period". Notes are per day, so a five-day period can hold five notes. Fine for bleeding, awkward for "cramps on day two".
- No stable period IDs for other features to reference. Anything downstream — predictions, stats, reminders — recomputes from the set.
- Bulk edits are several taps. A range-select gesture would be faster; we have not built it.
These are real costs. They were still cheaper than keeping two representations of the same truth in sync.
Why it matters for the privacy story
When the stored shape is a set of dates, three things get simple:
- Export is a list of dates a human can read.
- Delete means clearing a set, not cascading through period entities, caches and foreign keys.
- There are no server-side IDs to sync, because there is no server unless you opt in.
Derived values — cycle length, averages, prediction windows — are computed at read time from that set. Nothing derived is persisted, so nothing derived can go stale or survive a deletion.
The boring part is the point
A calendar that behaves like a data structure lets people mis-tap, change their mind, and correct a record months later without reading a help article. You never ask them to translate their life into a form.
The app I work on is Mori, a period tracker for Android that keeps records on the phone. No accounts, works offline: https://www.moricycle.com
If you are building anything where people record events over time, ask whether you need an entity or just a set of days. The entity is usually a UI idea wearing a data model's clothes.
Written with AI assistance.
