# Analysis review - 10 Nov 2026

Source of truth: `03_review_checks.py` / `.out.txt`. Cohort = users with a full 24h of follow-up (install_local + 24h <= end of 8 Nov local), n = 1,502.

## Fixed
- **Timezone bug.** DuckDB read `install_ts_utc` as TIMESTAMPTZ and cast it to Prague time, shifting installs 1-2h late (24h window became 25-26h). All scripts now `SET TimeZone='UTC'`. After fix: 0 events before install, 1,545/1,545 first opens match install time.
- **Cohort cut.** One rule in all scripts (was 1,470 vs 1,507).
- **Overlapping outcomes.** Granted/denied/skipped summed >100% because deniers then skip. Now an exclusive split.

## Retracted
- ~~"1.1.0 raised the grant rate 59% -> 66%"~~. Mix artefact: paid social test was all 1.0.0. Non-paid users: 51.5% vs 51.4% granted (p = 0.62). 1.1.0 did nothing measurable for activation.

## Holds (verified)
| Claim | Number | Check |
|---|---|---|
| Activation | 33.2% | `block_held` always has a matching `shield_shown` 5 min earlier (31,046/31,046) |
| Permission is the gate | 46.5% grant in 24h; 0 shields for never-granters | act given grant 71%, held given shield 95% |
| Denial is near-final | 429 deniers, 3% ever grant, 2.6% ever activate | re-ask exists (98 in onboarding, 15 from settings) but rarely used |
| 25% never reach permission screen | 374 users | 7 screens, 5-8% lost per screen |
| Paid social test drags the average | 20.2% vs 36.9% | z = -5.7, p < 1e-7; 55% of 19 Oct week |
| Weekend + work-hours schedule = no first block | 09:00 Mon-Fri group installed Fri-Sun: 38% vs 71% Mon-Thu | all 421 09:00 groups are days=12345 |
| Dead-end session tap | 67/67 who chose "start session" without permission were refused; 1.5% ever grant | plus 79 from home |

## Limitations
- Synthetic data; correlations, not causes. Fix effects need an A/B test.
- Can't see why users deny (no screen-level dwell or copy variant).
- Late installs have <24h, excluded; 1 Nov DST already handled in local ts.
