What are we working on?
Leave this tab and Jack starts to rot, live. Come back and you have to confess where you went.
Focusing on
Untitled grind
- Trips away
- 0
- Unexcused
- 0s
- Work passes
- 2
Confession time
You were gone 0s.
Pending damage: +0% rot
⚠️ This one is fatal unless it was work.
Where'd you go? Be honest. Jack can tell. (Jack cannot tell.)
Break time
Go scroll. It's legal now.
Jack is napping. Leaving the tab won't rot anything until the break ends.
How it works
- 1Start a session. Jack watches this tab while the timer runs.
- 2Wander off and Jack rots in real time. Mold, flies, the works. Even the tab icon rots.
- 3Come back and confess. Real work in another tab gets excused (twice). Finish and Jack heals and grows.
Your patch
Focus stats and a full record of what's been killing your pumpkins. Stored only in this browser.
Pumpkin killers
Time away by where you said you went, across every pumpkin.
Jack
- Focus
- 0m
- Sessions
- 0
- Fresh
- 100%
Recent sessions
| Date | Task | Length | Trips | Grade |
|---|
Insights
Your focus, modeled. Every number on this page is computed in this browser from your own session log. Nothing is uploaded.
How long you last
Share of sessions still slip-free at each minute (Kaplan–Meier). The band is a 95% interval. Sessions that ended clean are censored, not counted as failures.
When you slip
Slips per focus hour at each hour of the day. Faded bars have under an hour of data, so they're pulled toward your average (empirical Bayes) and one bad night can't define an hour.
What raises your risk
How each factor changes the odds of slipping in a given minute, from a logistic hazard model trained in your browser. Lines are 95% intervals. Faded bars aren't settled yet.
What each distraction costs
Every confessed trip away, grouped by where you said you went.
| Where | Trips | Typical | Worst 10% | Total |
|---|
Does the model work?
The model is trained on your oldest 70% of sessions, then scored minute by minute on the newest 30%, which it has never seen.
Each dot is a fifth of the test minutes: predicted risk (across) against the share that really slipped (up). Near the line means well calibrated.
Do Jack's warnings work?
When the model rates the next 5 minutes as risky, Jack warns you half the time, picked at random. The other half is the control group. This is a micro-randomized trial: comparing the two measures what the warning itself does, not just a correlation.
The data
Gourdian keeps a minute-level log of each session in this browser: when it started, when you left the tab, for how long, and where you said you went. No account, no server, no analytics scripts. Download it, or delete it from the Stats tab.
The graveyard
Every pumpkin you've lost. They watched so you didn't have to.