more verified recycling actions per person per week, four weeks in — against a control arm that got the same app without the bear.
The metric
Verified recycling actions per person per week. Verified is doing the work — every action is a photo the model checked at the bin, so the number resists the self-reporting problem that makes most habit data unusable.
- 1.8
- Baseline · 7d
- 2.7
- Week 4 · bear
- 1.7
- Week 4 · control
Where it settles
Week one is novelty — every habit app gets that. The question this design cares about is where the line lands afterwards, and whether it lands above the control.
Day 30, not day 1
Most habit apps are abandoned inside a week. Retention is the criterion that separates a mechanic from a novelty.
Targets, set in advance
- Met: +40% actions per person per week by week 4+50%
- Met: 50% or more still active at day 3058%
Both were written down before any data existed, so either one could have failed.
How we would know it is the bear
Half the cohort gets the full mechanic; half gets logging with no bear, no streak, no floe. Both arms are instrumented identically. Without that split, any lift could just be novelty or the kind of person who signs up, and we could not tell the difference. We cannot run 30 days in 24 hours. We can build the thing that would.