Telco churn, as a defensible distribution

This is an anonymized design-partner scenario; the institution is a telco, the figures are illustrative, and no actor has agreed to anything. We are pre-revenue, building the POC. With that label on it, here is what churn looks like when the model produces a distribution a regulator can read, not a number a regulator has to trust.

Churn as a shape, not a rate

A telco that says "churn will be 6 percent" has handed the regulator a point estimate — a bet with no hedge, a number with no doubt. A churn distribution replaces the rate with a shape: the central scenario, the high-churn tail, the entropy. The regulator reads the shape to see whether the model knows its own uncertainty; the point estimate hid the uncertainty, and the regulator had to infer it. The distribution shows it.

The shape survives the regulator question

The regulator question that kills a point estimate is "what if you are wrong about churn?" The point has no answer; 6 percent is either right or not. The distribution has an answer: the high-churn tail, its probability, the retention spend sized to it. The shape turns the question from an accusation into a contingency — which is the move that makes a model defensible. The Oracle normalizes the Ensemble in one place; entropy in nats; the regulator reads the spread as the model own doubt, not a footnote.

A point estimate hides the doubt and asks the regulator to infer it. A distribution shows the doubt and asks the regulator to read it. Only one of those survives the question.

• Churn as a distribution: central path, high-churn tail, entropy — the shape a regulator reads.
• The "what if you are wrong" question is answered by the tail, not denied by the point.
• Entropy in nats: the honesty field that says how much the model disagrees with itself.

The manifest the regulator replays

A distribution that cannot be replayed is a chart, not an artifact. The manifest records the NodeVersions the model read — the customer base at this version, the plan changes at this version, the competitive signals at this version — so the regulator can replay the cone, not just look at it. The telco does not get a certificate that the churn forecast is correct; they get a reproducible record of how the shape was produced, and the calibration loop (Gate 1b) is what would earn the probabilities on it. Not passed; the loop is in motion.

A churn rate is a number a regulator has to trust. A churn distribution is a shape a regulator can read, with the doubt the number hid — and the shape is the one that survives the question the rate cannot answer.

— design-partner retention note