Insights without a model

The apps PaceStreak gets compared with sell “AI feedback”: send your day to a language model, get back a paragraph. We have two rules that make that a non-starter. Nothing you log goes to a third party, and nothing on a screen may claim something the data doesn’t support. A model writing confident sentences about your sleep breaks both.

So insights are arithmetic.

What it compares

Each insight pairs a driver with an outcome: sleep hours with the share of habits kept, the morning energy rating with whether you trained, training with mood, protein with mood, and a few more. There are thirteen pairs, each chosen because it reads as a sentence someone would act on.

For each pair, the engine takes the last 90 days and splits them in two. For a yes/no driver like training, that’s days you trained against days you didn’t. For a number like sleep, it’s at or above your median against below it. Then it compares the outcome’s average across the two halves.

When it stays quiet

Most of the code decides when not to say anything:

• Both halves need at least 8 days. Three good nights prove nothing.
• The gap has to matter: 10 percentage points of habits, 0.4 on a 1-5 mood, 10% of what you usually eat.
• A Welch t-test has to clear |t| ≥ 2: roughly, the gap is unlikely to be noise given how much the days vary.
• Only the strongest insight per outcome is kept. Three ways of saying “you keep more habits when…” is one finding.

With a fresh account, that means an empty screen that says what to log to see more. That’s the correct answer. It isn’t a bug.

Correlation, said out loud

“You kept 92% of your habits after 7 h+ of sleep, against 42% on shorter nights” doesn’t mean sleep caused it. A calm week could cause both. The app says so under every list of insights, and it calls them hints worth testing, not findings.

Why not just use a model?

A model is better at prose. It’s worse at the one thing this feature is for, which is being right about you. The t-test is fifty lines anyone can read in app/insights/engine.py. Every number it shows can be checked against your export.