The Contextual Notification Sieve

The problem was never a single ping. It’s the volume forming behind it — and the fact that each app decides on its own when to interrupt you, optimising for its engagement, not your attention.

One promo is easy to ignore. Eighty notifications across a working afternoon is not. Sieve sits in front of the stream and runs one loop on everything: intercept → distill → escalate. It reads your mode, calendar, and location, silences the engagement bait, holds what can wait for a calm digest, and surfaces only what crosses the bar. The first design decision was to make context the arbiter — not each app’s notion of “important” — so on a focused day roughly 91% of the noise never interrupts you at all, and the decision behind every one is inspectable.

The signature move: take a cryptic, machine-voiced ping and turn it into one plain-language decision, the few facts you need, three prepared actions, and the reasoning behind the interruption — all on one screen.

The raw alert just said “Flight Update: UA2419.” Sieve read it, cross-referenced your calendar and booked connection, and worked out what it means: a 45-minute delay puts you at the gate after boarding closes. It states the decision, lays out land / connect / short-by as a fact grid, and prepares three one-tap actions — each previewing its full effect before anything happens. Then it opens its reasoning: an 88% confidence, the four signals it used, and a note on what it held in the same window. Why this is the best choice: an escalation people can’t verify is just a louder notification; legibility is what earns the right to interrupt. The benefit: you act in seconds, and nothing was booked on your behalf — the final call stays yours.

A third-party app can read Apple Health / HealthKit with permission, so this is a channel Sieve can legitimately ingest. The discipline is in how: amber not red, a pattern not a single reading, and a flag to look at — never a diagnosis.

The watch’s own alert just said “Irregular rhythm detected.” On its own Sieve would have held that — one reading is noisy. But this is the second flag in 8 days, both at rest, with resting heart rate drifting up from your own baseline — a pattern over time, and that’s what crosses the bar. Two telling design choices: confidence is shown lower (76%) on purpose, because honest uncertainty matters most here; and Sieve is deliberately quicker to surface, because missing a real cardiac pattern is the kind of mistake that’s hard to take back. Why this is the best choice: health is where overclaiming does real harm, so the product leans into calm and routes to actual care. The benefit: it brings forward the one pattern worth a look, holds 80+ everyday nudges, and never pretends to be a doctor.

The obvious calls Sieve makes silently. Review surfaces only the genuinely close ones — and crucially, it shows both kinds of mistake: something it held that maybe shouldn’t have, and something it interrupted you for that maybe could have waited.

A text from Sam with a “need to know tonight” deadline got held because Sam isn’t in your inner circle — Sieve is only 58% sure it read that right. A wishlist item came back in stock and Sieve pinged you, but nothing about it was time-critical — only 52% sure it was worth the interrupt. Each gets a one-line confirm or correct, and the lesson is shown already in effect. Why this is the best choice: most systems only ever expose false holds (things you missed); surfacing false surfaces too is what makes the bar honestly yours. The benefit: owning the close calls — in both directions — is how Sieve earns the right to keep deciding.

Cold-start is where most “smart” products quietly fail — they pretend to know you on day one. Sieve does the opposite: it says it doesn’t know your patterns yet, leans toward showing a little more, and asks when it’s unsure.

Now opens on “Day one together,” a calibration meter that reads just starting, and an honest banner: “I’ll show a little more and ask when I’m unsure this week — then quiet down.” Underneath sits the day-one prior it leans on before learning anything — your inner circle, security, and time-critical travel always reach you — plus a couple of items it’s openly unsure about, each with a one-tap “worth showing you?” rating. Why this is the best choice: trust is won fastest by being upfront about what you don’t yet know. The benefit: two taps today save a hundred interruptions next month — and it’s the only time Sieve asks for more than it gives.