Referent drift
Referent drift is what happens when a metric keeps computing exactly what it always computed, but what the team understands the number to be about has quietly changed. The dashboard looks the same and the count moves the same way, so nothing in the math flags that the meaning behind it has shifted.
Definition
The failure mode where a metric's referent goes fictional while the metric continues computing correctly.
How metric and referent come apart
The metric is what gets computed — the number a calculation produces, exactly, regardless of how anyone interprets it. The referent is what the team takes the metric to be about: the real-world phenomenon the number is read as describing. These can come apart in two structurally distinct ways.
For a singular metric tracking a multi-intent user base, "users approaching the resulting state" has coherent meaning when one intent dominates and loses it when sought states diversify. The team is reading the number against a unified referent that no longer exists; the metric keeps aggregating, but what it aggregates over has fragmented into multiple referents the singular metric can't address.
For an individual value milestone, the fiction creeps in as conditions shift. The milestone keeps firing, the count keeps moving, but the intent the team theorized as anchoring the correlation has drifted away from what's actually driving the firings. Same number, different population beneath it.
Catching referent drift
The fiction lives in the team's interpretation, not in the computation. The metric continues computing exactly what it computes — that's what makes referent drift hard to notice. There's no metric-level rigor that prevents it; referent drift operates at the interpretive layer while the computation proceeds untouched.
Detection has to operate at the same level as the failure: a maintained theory of why the metric correlates with the outcome, kept alongside the number itself, so the team can notice when the theorized referent has come apart from the empirical firings. A portfolio of milestones makes per-intent referent drift diagnosable rather than blended into an averaged signal. Both are instances of model-level reasoning operating where metric-level reasoning cannot.
Example
A project management tool launches with a milestone of "first task assigned to a teammate," anchored to the team-coordination intent that dominates its early signup cohort. The milestone correlates well with first paid conversion. Two years in, marketing has shifted toward solo founders and freelancers; many new signups now assign tasks to themselves as personal reminders. The milestone keeps firing at a healthy rate. The team reads the number as "team-coordination intent landing successfully" — that's still what the dashboard label says, that's still how it's interpreted in weekly review. But what's now driving most firings is solo self-organization, a different intent the milestone was never theorized against. The metric computes correctly; the referent has gone fictional.