Documentation · the method

The science behind the engine.

Brume applies the just-in-time adaptive intervention (JITAI) paradigm. Where classic apps count days and react after the fact, Brume flips the logic: measure continuously, estimate risk windows, intervene before the peak. Here's how — and what the science says, or doesn't yet.

The discipline of numbers

Three families, never mixed.

Every number belongs to a family. We never conflate them — it's the project's rule of rigor.

Science

From a publication, always cited with its comparator, sample size (n) and reference.

Product

A design goal or promise, flagged with "~" or "target".

Demo

A fictional interface value — plausible, coherent, never presented as a result.

The problem

Relapse isn't a failure of willpower.

The urge to smoke isn't a continuous state: it's a series of brief episodes (~3 min), triggered by conditioned contexts — coffee, the end of a meal, the break, alcohol. Smoking is a ritualized behavior, structured by the clock of the day. That's exactly what makes it predictable.

~30

attempts on average before quitting for good. Relapse is the statistical norm, not a fault — Chaiton et al., 2016.

The paradigm

Just-in-time interventions.

A JITAI delivers the right support, at the right moment, only when the person needs it and can receive it (Nahum-Shani et al., 2018). Four components, mapped onto Brume:

Tailoring variables
The 6 check-in variables, time of day, the FTND profile, your triggers.
Decision points
The estimated risk windows — plus any manual trigger (SOS).
Intervention options
60-second exercises: urge surfing, breathing, redirection.
Decision rules
The hybrid engine: clinical rules as foundation, personal model on top.
16.4% vs 10%

The Smart-T trial (RCT, n=454, JAMA Netw Open 2025): verified abstinence at 26 weeks of 16.4% vs 10% for a standard app. It's the only smartphone JITAI validated for efficacy — Brume's scientific archetype.

The architecture

Measure → Predict → Intervene.

A closed loop running at the scale of the day. Every episode outcome — urge ridden out or relapse — refines your windows.

01

Measure

Micro check-ins (EMA) of 6 questions, ~20 s. The 6 validated variables of Smart-T's risk score.

02

Predict

A hybrid engine, chronobiology-first. Time predicts better than place: 87.9% F1 across 1,784 episodes.

03

Intervene

A nudge ~20 min before the estimated window. A brief 60-second intervention, before the peak — not during.

The 6 measured variables
  • Craving intensity
  • Stress level
  • Alcohol intake
  • Smokers around you
  • Motivation to stay abstinent
  • Cigarettes within reach
Across 24 hoursEstimated risk
6h12h18h0hCoffeeBreakEvening

The dominant signal is time, not place. Removing GPS barely degrades the model; removing temporal features collapses it (npj Digital Medicine, 2025).

The safety net: an independent SOS.

Since the engine will miss episodes (AUC 0.63), an SOS urge button is always available — bypassing prediction, with no window condition. It stays free for life. Prediction helps you anticipate; it's never a single point of failure.

What the science doesn't (yet) say

Our limits, in the open.

Scientific honesty is a trust asset, not a weakness. Three limits, publicly owned.

Fine-grained prediction stays modest

Forecasting an episode an hour ahead with ML alone tops out at AUC ≈ 0.63 — barely better than chance. It's the field's structural limit.

An app isn't a treatment

Nicotine replacement and varenicline remain the reference treatments. Brume always points to a professional and a quitline.

The companion supports engagement

The avatar boosts content use (+30%), not abstinence directly. We'll never claim otherwise.

The evidence base

Every choice, a study.

Paradigm archetype
Smart-T · JAMA Netw Open 2025 · RCT n=454
16.4% vs 10% at 26 wks
Only JITAI validated for efficacy
Engagement as mechanism
iCanQuit · JMIR 2022 · n=2,415
56% vs 23% (×5 odds)
Association, not causation
Dominance of the temporal prior
npj Digital Medicine 2025 · 1,784 episodes
87.9% F1 · GPS marginal
Classification of observed episodes
Limit → hybrid architecture
PLOS ONE 2026 · NTR 2023
AUC ≈ 0.63–0.66
The field's owned limit
Cautious use of context
Quit Sense · NTR 2023 · n=209
11.5% vs 2.9%
Positive but underpowered signal
Companion rationale
"Inner Dragon" · JMIR 2024 · n=479
+30% content use
Engagement effect only

Ready to see the craving coming?

Do the 60-second profile and join the first cohort. Free for the first sign-ups.

Brume is behavioral wellness support — not a medical device. Need help quitting? Reach out to a quitline.