# Retention + DAU cut (14 Nov 2026, validated)

Validation: `05_retention_checks.py` / `.out.txt` (data edges, fixed-cohort curve, channel/version confounders, weekday bias, 95% CIs). Corrections from validation are marked **[checked]**.

Script `04_retention_dau.py`, output `04_retention_dau.out.txt`. Cohort rule as `03_review_checks.py` (full 24h of follow-up, n = 1,502, activation 33.2%). Day n = local calendar day since install; a user counts for day n only if they have data through that day. Data ends 8 Nov, so the longest clean read is **D21 (n = 143 activated)**. D30 cannot be measured yet.

Three meanings of "active" for a blocker:
- **Opened**: opened Lull that day (classic DAU).
- **Protected**: at least one held block that day (the value Lull delivers).
- **Any**: any event, including block screens seen without opening Lull.

## 1. Activation is the gate to everything

| Day n retention | Activated (held <24h), n=499 | Not activated, n=1,003 |
|---|---|---|
| D1 any / protected | 79% / 68% | 6% / 1.5% |
| D7 any / protected | 52% / 43% | 2% / 1.2% |
| D14 any / protected | 34% / 26% | 0.4% / 0.3% |
| D21 any / protected | 27% / 15% | 0% / 0% |
| Week 2 (days 7-13) protected at least once, users with 13+ days | **50%** (n=324, CI 45-56) | **1.5%** (n=718, CI 0.9-2.7) |

Dn columns are read on the same weekday as install, so they carry a weekday effect (Mon-Fri schedules). **Week 2 is the headline metric** because it covers all 7 weekdays. **[checked]**

- **Almost all retained users are activated users.** A non-activated user is effectively gone by day 3.
- **It's not just "people who gave permission are keener".** Users who granted Screen Time in 24h but had no held block retain far worse than activated users: D7 any 7% vs 52%, protected 4% vs 43%.
- **Holds in every segment. [checked]** The week-2 gap holds inside every channel (activated 34-67% vs not 0-6%) and both versions. Paid social activators retain least (34% vs 42-67% for other channels), so the paid social test hurt both activation and the quality of the activators it brought.
- **The decay is real, not a cohort-mix artefact. [checked]** A fixed cohort (143 activated users with 21+ days) shows the same curve: protected 68% D1, 44% D7, 29% D14, 15% D21.
- **Late activation still works, but n is tiny.** 29 users had their first held block on day 1-6; 9 of them (31%, **95% CI 17-49%**) were protected in week 2, vs 50% for <24h activators and 0% for never. A second chance after day 0 is worth building, and testing.
- Correlation, not proof: intent drives both. Only an activation test proves the lift. That is why Fix 1 ships as an A/B test.

**Sizing Fix 1 in retained users.** Week-2 protected users per 100 installs today: 0.332 x 50.3 + 0.668 x 1.5 = **17.7**. If activation goes 33% -> 42% (+8.8 pts, itself an assumption), the extra activators' week-2 rate decides the gain: 17% (late-activator CI low) -> 19.1 (**+8%**); 31% (late-activator point) -> 20.3 (**+15%**); 50% (today's activators) -> 22.0 (**+24%**). For the deck: **+8% to +24% retained users, central +15%.** The killer assumption is how well the extra activators stick, so the test needs week-2 protected as a secondary metric, not just activation. **[checked]**

## 2. "DAU" undercounts and miscounts a blocker

Active user-days after install day (day >= 1):

| Kind of day | Share |
|---|---|
| Opened Lull and protected | 54% |
| **Protected, never opened Lull** | **23%** |
| Opened Lull, no held block | 22% |
| Block screens / breaks only | 1% |

- **23% of the value is invisible to opened-DAU.** Lull worked that day and the user never opened it. A good blocker *should* need less opening over time.
- **22% of opened days delivered no protection.** Opens alone overstate value.
- Over time the two curves cross: for activated users, opened falls faster than protected (D7: 38% opened vs 43% protected; D12: 23% vs 31%).
- **Protected users have a weekly pattern by design. [checked]** Averaged over 12 Oct - 8 Nov, protected-DAU is ~129 on weekdays vs ~107 at weekends (-17%), and up to -26% vs Thu/Fri. The whole dip comes from users with a Mon-Fri schedule (~60 on weekdays -> ~29 at weekends); everyone else is flat at ~78. Report **weekly protected users** (or a 7-day average), not raw daily.
- Per activated user over days 1-14: 5.6 opened days vs 5.7 protected days (correlation 0.83). The two are related, but they are not the same thing.

**North star proposal:** **Weekly Protected Users** (a held block in the last 7 days). Opened-DAU sits next to it, as the brief's DAU focus asks. Guardrails: break rate per block screen, groups switched off, sessions ended early, Screen Time disconnects.

## 3. Guardrails and what gets blocked

- 80% of block screens hold; **19% end in a 5-minute break**. The break rate is flat across apps (17-22%), so it is a habit, not an app problem.
- Strict groups hold 100% of block screens but are only 7% of all block screens. Strict mode is a lever, but a niche one.
- Over the 30 days, 220 users switched a group off, 244 ended a session early and 31 disconnected Screen Time, out of 552 who ever had a held block.
- Most blocked apps: Instagram and TikTok (~4.4k held each), then WhatsApp (3.7k), Snapchat, Reddit, Facebook, X, YouTube. Telegram and Discord have the highest break rate (22%).

## 4. Goal matters for who stays

First pass used D7 protected, which is biased: every "focus" user has a Mon-Fri schedule, and D7 falls on their install weekday. Re-cut on week 2 (weekday-neutral), activated users with 13+ days: **[checked]**

| Goal | Users | Activation | Week-2 protected, if activated | With a Mon-Fri schedule |
|---|---|---|---|---|
| Less social media | 436 | 39% | **41%** (50/121) | 13% |
| Better sleep | 289 | 47% | **44%** (36/82) | 10% |
| Focus at work | 427 | 31% | 63% (51/81) | 100% |
| Time with family | 161 | 38% | 65% (26/40) | 8% |

- **"Social" and "sleep" users drop fastest** (41-44% vs 63-65%). The social vs focus gap is real (CIs ~33-50% vs ~52-73%); family has only n=40. Social users fight Instagram and TikTok all day rather than in a set window. This is where a positive loop (Unrot-style earn-your-time) or a narrow scope (SocialLite-style) fits best.
- **"Focus" users activate worst (31%) but stay best.** The weekend problem sits here: a Mon-Fri 09:00-17:00 schedule set up on a Friday doesn't fire until Monday.
- Direction, not precision: 40-121 activated users per goal.

## So what for the deck
1. **Activation is the lever** (Fix 1). 50% vs 1.5% week-2 retention is the slide.
2. **Success metric = Weekly Protected Users**, with opened-DAU shown next to it, and the reason why: 23% of protected days have no open.
3. **Second problem after activation: protection decays** (fixed cohort: 44% D7 -> 15% D21 protected). The improvements beyond blocking should target this, starting with "social" and "sleep" goal users.
4. **Monetisation trade-off:** revenue can only come from the ~1 in 6 installs who are still protected in week 2. A paywall before activation taxes the gate, so the paywall must come after the first held block.

## Limitations
- Correlational: activation vs retention is not causal. Only the Fix 1 A/B test can show the lift.
- D21 rests on 143 activated users from the first 9 days of installs. D30 is not measurable.
- "Protected but never opened" days: 94% of their held blocks come from groups (schedules), 6% from focus sessions. Sessions starting without an open on that local day are probably sessions that ran past midnight. Minor, but not verified.
- Synthetic data (per the data dictionary). Patterns may be cleaner than real life.
