Rules — Review queue

What is a rule?

A rule is an additive score adjustment triggered by a condition on a signal. Example: "if a trial has 10+ days without login, subtract 20 pts". Rules extend the 5-pillar formula (activation, engagement, depth, team, momentum) without rewriting it — they capture patterns the pillars miss.

Where do candidates come from?

Every Monday at 05:30 Paris, the rule miner brute-forces thousands of combinations of (signal × threshold × cohort × outcome) across 30 days of training data. For each one, it asks: "does the outcome actually happen at a different rate when this condition fires, or could it just be luck?" (the technical test is a two-proportion Z-test). Because we try thousands of patterns at once, a few will look "significant" by pure chance — so we apply a correction that caps the expected false-positive rate among kept patterns at ~10% (Benjamini-Hochberg FDR, q=0.10). The 20 strongest survivors (ranked by effect size × sample strength) get surfaced here. Tautological patterns (e.g. "logged in today → less churn" — already in the engagement pillar) are filtered upstream.

What can you do with a candidate?

• Accept — pattern looks genuine and actionable. Moves the rule to the Accepted pool (still NOT active — no live score impact yet).
• Attach to formula — pick a candidate formula (shadow), copy the rule into its definition.rules[]. The rule then runs in shadow for 14 days; if the candidate beats the active formula, promotion to live is a manual migration.
• Reject with a note — pattern is tautological, confounded, already covered by a pillar, or domain-implausible. The reject note is kept to help future mining runs learn.

Column cheatsheet

• Support — how many (location, day) rows matched the condition during the training window. Think of it as the sample size. We require ≥ 30: with fewer observations, any difference we see could easily be random noise rather than a real pattern.
• Lift — the gap between the outcome rate inside matching rows and outside (rate_in − rate_out), expressed in percentage points. Lift of +12 pp means the outcome happens 12 pp more often when the condition is true than when it's false. Whether that's good or bad depends on the outcome — a positive lift on "churned" is bad news; on "converted" it's good news.
• p-value — how likely it is that this pattern showed up by pure coincidence, if the condition actually had no real effect. 0.01 means "if there were truly no signal, we'd see something this strong about 1 time in 100". We only surface patterns under 0.05 (max 1 in 20 could be a fluke) and additionally pass a multiple-testing correction (BH-FDR q=0.10) so that ≤ 10% of what we show you is expected to be noise. Lower p = less likely to be coincidence.
• Δ — the score adjustment this rule would add (or subtract) if attached, in points (−30 to +30). Signed so that a "bad" outcome lift becomes a negative delta (penalty), and a "good" outcome lift becomes a positive delta (bonus). It's a starting suggestion — you can tune it before attaching.