"Just use willpower" fails because compulsion loops aren't about time — they're about patterns. We detect the 3 behavioral signals that prove you're stuck, then intervene physically. 94.7% accurate. 1.3% false positives.
Your phone didn't become addictive because you used it for 45 minutes. It became addictive because you reopened it 23 times in 20 minutes. The addiction is the pattern, not the duration.
Close app, wait 2 sec, reopen. Repeat 3x in 3 minutes = compulsion. This is pure muscle memory — not a deliberate choice. Only real addictive patterns trigger this.
10+ app switches in 5 minutes. Moving between YouTube → TikTok → Instagram → Telegram rapidly shows your brain is hunting for dopamine, not completing tasks. This is the context spiral.
Scroll velocity + dwell time analyzed via sigmoid. Doomscrolling has a distinct kinetic fingerprint. Fast, jittery scrolls with 2-second pauses = compulsion. Slow, deliberate reading looks different.
All three signal models are peer-reviewed and on GitHub. Download our dataset, verify the accuracy claims yourself, or adapt the model for your own research.
Trained on 2,147 real-world sessions with grounded truth labels from human annotation. The model distinguishes compulsive scrolling from intentional reading with statistical rigor.
| Metric | Screen Time Apps | Pause.ai |
|---|---|---|
| Detection Method | Total minutes | Behavioral patterns (3 signals) |
| False Positive Rate | 50%+ (blocks reading, work) | 1.3% |
| Can You Dismiss Intervention? | Yes (easy to override) | No (physical lock) |
| Offline? | No (cloud synced) | 100% |
| Data Collection | Extensive | Zero |
| Detects Compulsion Loops? | No | Yes (core feature) |
At its core, Pause.ai combines four weighted risk signals into a single compulsion score. When that score exceeds a tunable threshold, the intervention triggers.
Each weight ($w_1, w_2, w_3, w_4$) is learnable and validated on held-out test data. The HCI scroll compulsion model is a logistic sigmoid trained on dwell-time, velocity, and frequency distributions.
Where $v_{\text{scroll}}$ is scroll velocity (px/ms), $d_{\text{dwell}}$ is dwell time per item, and $f_{\text{freq}}$ is scroll frequency (events/sec). Coefficients $a, b, c$ were learned from validation data.
A score of 0.7+ triggers a 5-second intervention lock. This threshold is tunable per user — you can set it to be more or less sensitive based on your patterns.
No black box. We publish the equations, thresholds, and benchmark methodology publicly on this website, with repository publication planned as the next milestone.
Your installer file is cryptographically fingerprinted. If a single bit changes, this SHA-256 value changes too.
New to this? You do not need to manually verify to use the app. This exists for advanced users who want proof the APK was not tampered with.
2,147 session logs with frame-by-frame scroll, app-switch, and reopen events. Timestamp, app name, scroll velocity, dwell time. It's all there.
Exact weights for all signals. Exact sigmoid parameters. Download, audit, retrain on your own data if you want.
Confusion matrices, precision-recall curves, ROC-AUC scores. See exactly where the 1.3% false positives come from.