# Protection state + why protection fades (14 Nov 2026)

Script `06_protection_state.py`, output `.out.txt`. Answers two points from the 14 Nov second opinions: (A) "held block that day" penalises success; (B) diagnose the fade before choosing retention fixes.

**Under protection (eligible)** on a local day = Screen Time granted and not disconnected, and at least one standing group switched on that covers that weekday (schedule days, or limit/always). Focus sessions are excluded. Reconstructed from group_created / edited / toggled and permission events. Sanity: 8.5% of held-block days fall outside "eligible" (sessions and approximation), so it is close but not exact.

## A. "Under protection" can't be measured cleanly from this data
- Users under protection: ~400-550 a day (last 2 weeks), vs ~115-175 with a held block. Only 31-40% of under-protection days have a held block, falling over time.
- **But the base is inflated by deleted apps.** Deleting Lull removes its blocks and fires no event, so a deleted user still looks "under protection". On under-protection days after install, 62% have no event at all.
- **So the reviewers are right in principle, and the data can't settle it.** Recommendation: keep **protected users (held block)** as the measurable north star now, and **add a daily heartbeat event** (e.g. from the Screen Time monitor extension) so "under active protection" becomes measurable. That is an instrumentation fix worth a line in the deck.

## B. Why protection fades: activated users with 14+ days follow-up (n = 310), state on day 14

| State on day 14 | Users | Share |
|---|---|---|
| Protected (held block) | 74 | 24% |
| **Silent: blocks on, no event at all** | 137 | **44%** |
| **Switched all groups off** | 62 | **20%** |
| Screen Time disconnected | 19 | 6% |
| Groups on but none covers day 14 (weekend) | 8 | 3% |
| Saw a block screen and got round it | 3 | 1% |
| Opened Lull, no block screen | 7 | 2% |

Day 7 for comparison: 41% protected, 34% silent, 14% groups off, 5% disconnected.

- **Silent is the biggest group, and it starts early.** Of the 137 silent users, **85 had their last event of any kind on day 0-3**. A user losing 3-7 hours a day to their phone, with blocks switched on, who triggers no block screen for 10+ days has most likely deleted Lull. It's an inference, because uninstalls aren't logged. That is ~27% of activated users gone within 3 days of their first held block.
- **One in five switches blocking off deliberately** by day 14 (20%), plus 6% who disconnect Screen Time. That's a quarter giving up on purpose, through the off switch, not the block screen.
- **Heavy break-takers fade.** Activated users who ended 30%+ of their week-1 block screens with a break: 31% protected in week 2, vs 57-59% for others (n = 55 vs 251; correlation).

## So what for the improvements
1. **The first 3 days after activation are a second cliff.** ~27% of activated users go silent by day 3. The first-week programme (day-1/2 nudge, "locked in", first weekly win) is the right improvement, more than a plant.
2. **The off switch is where people give up, not the block screen.** 20% switch all groups off; only ~1% are "got round a block" on a given day. A softer off ("pause until tomorrow" instead of off, ask why) targets the bigger leak than a pause on the block screen.
3. **Breaks are an early warning, not the leak.** 30%+ break-takers fade, so a break budget is a fair guardrail feature, but on its own it is not a top-3 improvement.
4. **Instrumentation gap:** add a heartbeat and an uninstall proxy so silent users can be told apart from not-tempted users.

Limits: synthetic data; uninstalls not observable; group state reconstructed; correlations, not causes.
