Pattern → Metric → Model
The Pattern → Metric → Model curve sorts teams into three levels by how they reason about user behavior. Pattern-level teams work from what things look like: competitor screenshots, UI conventions, design hunches. Metric-level teams work from behavioral data: what correlates with conversion, what thresholds users cross, what split tests show. Model-level teams incorporate an explanation for why those correlations hold, so the changes they ship target the underlying mechanism rather than the correlation alone. That explanatory layer, not extra rigor, is what separates model from metric.
Definition
A three-level curve of reasoning reliability about user transitions, distinguished by what each level treats as evidence: visual or comparative material at pattern-level, behavioral-statistical evidence at metric-level, and an explanatory framework for why proxies correlate at model-level.
The three levels
- Pattern-level operates on visual and comparative inputs — competitor analysis, UI polish, design pattern borrowing, executive intuition. The inputs don't connect causally to user behavior. Most teams live here.
- Metric-level operates on behavioral-statistical inputs — correlation with the business outcome, behavioral proxies, signup cohort splits, disciplined experimentation. A metric-level team can do rigorous work by its own standards.
- Model-level adds an explanatory framework for why proxies correlate. With that layer, the team can predict what would break the correlation, anticipate failure modes, and treat the correlation claim as ongoing rather than settled.
The explanatory layer
Craft quality doesn't separate metric from model. A metric-level team can be careful, statistically literate, and disciplined about experimentation. What it lacks is a theory of why the proxy it's optimizing correlates with the commercial outcome — which is what lets a model-level team detect when the correlation is starting to dissolve and what's causing it.
The canonical metric-level failure is post-definition silence on the correlation claim. The team establishes a proxy by demonstrating its correlation with a business outcome, then optimizes against the proxy for quarters without re-checking whether the work is moving the outcome the proxy was justified by. Model-level teams treat the correlation as a live claim: when the proxy moves, the team checks whether the business outcome moved with it; when the proxy doesn't move, the team checks whether the business outcome moved anyway.
What the explanatory layer enables
Two main advantages over metric-level work:
- Earlier canary on correlation drift. A model-level team detects when a value milestone's correlation begins to shift before the business outcome resolves the question through declining numbers. A metric-level team's first signal is the metric breaking, by which point the team is already behind.
- More reliable interventions. Model-level teams ground interventions in theorized mechanism rather than observed correlation alone. Interventions are less likely to ride correlation that wasn't carrying the proxy's value and more likely to address the structural mechanism the proxy was tracking.
Teams without the explanatory layer often don't experience drift as drift. They read metric noise, intervention churn, and performance variance as separate problems requiring separate fixes. The drift gets absorbed into those other categories, which is why the question "are you experiencing unexplained drift?" tends to come back "no" even from teams suffering from it.
What moves teams up
Not intellectual attraction to rigor. Commercial pressure — failed tests, wasted sprints, flat performance against the business outcome. The forcing function is economic, not philosophical. Teams move up when their current level stops producing the results they're being measured on.
How to diagnose where a team sits
A model-level team can speak to why a behavior correlates (not just that it does), predict what would break the correlation, generate hypotheses about what other behaviors should correlate, and form a view on whether the correlation would hold under a different acquisition mix. A metric-level team answers from observation rather than explanation and is regularly surprised by failure modes. These are signals, not proofs — full diagnosis usually requires waiting for scale to surface correlation drift.
Example
Twitter's "30 followers" sat in folklore as a great activation metric: users who organically followed 30 accounts retained well. A metric-level team treats this as a finding and reaches for interventions that produce the count — preload 30 celebrity accounts on signup with a "select all" checkbox. The count moves. Whether the business outcome moved with it is a question that doesn't get asked sharply, because the proxy is the target.
A model-level team would hold a theory of why organic following of 30 accounts correlated with retention — the user had assembled a feed worth coming back to. With that theory in place, the team predicts that manufacturing the count via celebrity preload won't reproduce the feed-assembly the original correlation reflected. The intervention either doesn't ship in that form, or it ships and the team checks the business outcome rather than just the count. The metric-level version absorbs each failure into local explanations — measurement, attribution, seasonality, the proxy needing to be paired with another driver — and the underlying theory never gets registered as wrong. The model-level version reads the failure as evidence about the proxy.