Notification anxiety and interruption residue, read from your own timestamps

The notification is the platform's handshake: it pulls you in, and the pull
costs more than the glance. That is the one part of the loop that never lands
in an export - push banners, badge counts and sound are delivery-side events,
recorded on your phone, not handed back to you. What does land is the response
side: every like, search, message and session-shaped cluster of activity with a
timestamp attached.

This post is about that honest seam - what the psychology of interruption
predicts, what the archive actually carries, and how to read one without
pretending it carries the other.

What the research says interruption does

Three findings are relevant, and all three are about switching, not about
duration:

• Attention residue. Mark, Gonzalez and Harris found that after working on

  Task A and switching to Task B, part of attention remains stuck on A, and the
  residue is stronger when A was unbounded or cognitively heavy.¹
  Checking a feed between work tasks does not cost the ten seconds of the
  check; it costs the ramp back into the task, repeatedly.

• The checking loop. The "elastic limit" model of smartphone checking

  describes a self-interrupting loop: a goal or doubt triggers a check, the
  check resolves (or fails to resolve) the itch, and the interval until the
  next check shortens when resolution is uncertain.² Uncertainty - not
  boredom - is the accelerator. That is why the unread badge is engineered
  vagueness.

• Cost of self-interruption. Workplace studies find people interrupt

  themselves almost as often as they are interrupted externally, and that
  self-interruptions cluster: once the checking habit opens a door, a sequence
  follows.³ A run of ten checks is behaviourally one interruption
  event, not ten deliberate decisions.

None of these studies needs the notification itself to matter - they are about
what switching does to attention. Which means the response timestamps, if
honest and complete, are fair evidence of the checking rhythm - just not of the
anxiety attached to it.

What your archive actually carries

The activity side of a typical export is a pile of timestamped rows: likes,
reactions, comments, searches, story interactions, follow events - each row
carrying at most a Unix-second instant.⁴ From those rows the product
builds, in-browser:

• The day-hour heatmap - a 7x24 grid over all your own actions, with 27

  narrower modes (likes only, searches only, messages only, and so on), every
  cell clickable down to the rows behind it.⁵

• Hourly rhythm and by-weekday charts - the same data as two distributions

  so a 2am cluster is visible as a shape, not just a dark cell.⁶

• Per-kind counts with dated/undated separated - rows without a parseable

  timestamp stay in totals and stay out of the grid, because an undated row is
  unknown, not zero.⁷

What the archive does not carry, and therefore what the app never implies:

• No notification records. You cannot see what you were shown, when the

  banner appeared, or whether you were pulled in versus already opening the
  app. A cluster in the grid is your actions, not the prompts that preceded
  them.

• No session boundaries. Rows are events, not sessions. "Checked 40 times"

  is not derivable - "made 40 timestamped actions" is. The gap between those
  two sentences is where honesty lives.

• No focus, mood or physiology. Nothing in a JSON timestamp says whether

  you were anxious, in flow, avoiding work or in bed with insomnia.

Reading a checking rhythm without self-prosecuting yourself

If you look at your own grid, these rules keep the exercise useful instead of
shaming:

1. Look at runs, not points. Consecutive evening cells with tight spacing

   describe a checking pattern - per the self-interruption research, likely
   one behaviour expressed repeatedly - rather than a series of independent
   choices you should judge individually.

2. Separate kinds before drawing conclusions. A late-night cluster of

   searches reads differently from a late-night cluster of story views, and
   the modes exist precisely so you never have to sum them into one number
   that means everything.⁵

3. Check the dated subset. Undated rows are excluded from the grid; totals

   and cells can legitimately differ, and a difference is not a bug.⁷

4. Ask the better question. Instead of "am I checking too much?" (a

   judgment the file cannot support), ask "what time does checking usually
   start, and what kind of action opens the run?" - both answerable from the
   rows themselves.

The point of the exercise is agency: seeing the shape your own attention took
in hours you no longer remember, in a file that never leaves your device. That
is a different gesture than a phone's Screen Time dashboard - same
timestamps, but no target attached, no streak to break, and no score at the
bottom.

Related reading

The heatmap post for how any timestamped row becomes a day-hour cell (and the
null-is-not-zero rule); the activity inventory for every source file feeding
the grid; the shadow-diary post for the other unperformed layer - what you
looked up rather than what you did.

¹: Mark, G., Gonzalez, V. and Harris, J. - "No Task Left Behind?
Examining the Nature of Fragmented Work" (CHI 2005) - attention residue on
switching between tasks; cited as research framing.
²: Elhai, J. et al. and the anxiety-uncertainty checking literature -
the "elastic limit" / uncertainty-accelerated checking account of phone
checking; cited as background, not as a claim this product tests.
³: Mark, G. et al. - studies of self-interruption rates finding
frequent self-initiated switches, often in runs; cited as framing for why
timestamp clusters are read as patterns.
⁴: packages/shared/src/analytics/activity-taxonomy.ts - activity kinds
carry at most unix-second timestamps; types/activity.ts ActivityRecord is
{ value, href, timestamp, title }.
⁵: apps/web-next/src/components/ActivityHeatmap.tsx - 7x24 buildGrid
over collectActivityEvents, HEATMAP_MODES covering 27 non-all scopes,
cells link to their rows.
⁶: apps/web-next/src/views/Activity.tsx - "Hourly rhythm" / "By
weekday" segmented charts over the same events.
⁷: analytics/activity-taxonomy.ts - raw counts are never
timestamp-filtered; the dated count is tracked separately, so grids and totals
can legitimately differ.
⁸: packages/shared/src/ama/tone.ts - refusal contract for questions
about inner life; questions about how you felt return absence answers.