Skip to main content

​
Maturity Ramp

When a new offer goes live, its propensity posterior is wide — we don’t yet know its true acceptance rate. Showing it at full ranking confidence is reckless (we might over-spend on a dud) but suppressing it entirely is also wrong (we’ll never learn). The maturity ramp is the gating decision that controls cold-start exposure as evidence accumulates. Kaireon ships two maturity ramps. The default — Bayesian Confidence-Bound Maturity Ramp (BCB-MR) — is the better one and the rest of this page covers it. A legacy linear-count ramp is kept for back-compat with tenants who tuned around a fixed interaction threshold.

​
Why BCB-MR exists

Count-based maturity gates — “treat the model as immature until it has accumulated N responses” — are the conventional choice in next-best-action systems. The approach is intuitive but has two real failure modes:
  1. Low-volume offers are trapped. A campaign for 100 high-value enterprise customers will never accumulate a 5,000-response gate. The model is forever immature even when its posterior is already tight — useful signal is thrown away.
  2. High-volume volatile offers are prematurely “done”. An offer with thousands of trials at a seasonal volatile rate still has a wide posterior. A pure count gate declares it mature even when it deserves continued exploration.
The signal we actually want is posterior credible-interval width — “how much room is left for the true rate to be?”. An offer is mature when we can pin down its rate within an acceptable confidence band.

​
The algorithm

For each candidate offer at decision time:
  1. Aggregate the offer-scope ModelAdaptation posterior for this offer — its positives + negatives summed across all active models. The maturity ramp keys off the offer-scope Beta posterior specifically, not the full scope hierarchy that propensity scoring walks — maturity is a property of the offer, not of a channel/direction cell. (A paused offer-scope row bypasses the ramp entirely.)
  2. Compute the Wilson score interval for the Bernoulli proportion p̂ = positives / (positives + negatives) at the configured confidence level (default 95%, z = 1.96). Wilson 1927 is the canonical closed-form CI; well-behaved at boundaries (no blow-up at p̂ = 0 or 1, unlike the Wald interval).
  3. Let width = upper − lower. If width ≤ widthThreshold (default 0.20), the offer is mature → exposureProbability = 1.0.
  4. Otherwise the offer is immature. Compute a decaying cold-start floor:
    At n=0 the floor is baseFloor (default 0.50). At n=100 with default decay it’s ≈ 0.15. The decay mirrors UCB1’s √(2 ln t / n) shrinkage but applied to the floor rather than an exploration bonus.
  5. Return exposureProbability = max(floor(n), wilsonLower). The max ensures we never penalize evidence we already have — an offer with strong early positives (8/10 → Wilson lower ≈ 0.49) gets at least that lower-bound exposure even if the decay-floor shrank below it.
The caller then does a deterministic per-customer/per-day roll: hash(customerId + offerId + today) ≤ exposureProbability decides whether THIS impression includes the offer. Observability. A ramp exclusion is never silent: every withheld candidate is recorded in the request’s rejectedOffers[] (stage maturity_ramp, reason "Below cold-start exposure floor — maturity ramp withheld this offer for this customer today (exposure probability N%, mode …)") when the caller sets explain=true, and on the persisted DecisionTrace.rejectedOffers for sampled traces. See Recommend and Decision Traces.

​
Worked examples

The table below shows real exposure decisions at default config (widthThreshold = 0.20, baseFloor = 0.50, z = 1.96, decayHalfLife = 10). The 8/10 row is the interesting one: the offer is clearly persuasive (μ̂ = 0.80) but the CI is still wide, so it’s ci_gated — exposure is the Wilson lower bound 0.49. As soon as evidence tightens the CI below 0.20 width, the offer flips to mature.

​
Configuration

Per-tenant settings (Settings → Models → Maturity Ramp):

​
Why this design works

  1. Low-volume offers aren’t trapped. A campaign for 100 enterprise customers can mature once its posterior tightens, regardless of raw response count.
  2. High-volume volatile offers stay in exploration. An offer with thousands of trials but a still-wide CI keeps getting exploration traffic instead of being declared “done” by a count threshold.
  3. The threshold is principled. A 0.20 width at 95% means “the true rate is known to within ±0.10” — a tunable that business owners can reason about, with a clear posterior interpretation rather than a chosen convention.
  4. The maturity event is observable. Operators see a single interpretable number — “this offer’s posterior CI width is 0.18, just below the 0.20 threshold” — instead of “evidence count is 4837 of 5000”. Combined with direction-scoped adaptations, this surfaces directly in the Model Health UI.

​
References

  • Wilson, E.B. (1927). “Probable inference, the law of succession, and statistical inference.” JASA 22:209–212. The canonical Bernoulli proportion CI.
  • Auer, Cesa-Bianchi, Fischer (2002). “Finite-time Analysis of the Multiarmed Bandit Problem.” Machine Learning 47:235–256. UCB1 — origin of the √(ln t / n) shrinkage form adapted here as a decaying floor.
  • Chapelle & Li (2011). “An Empirical Evaluation of Thompson Sampling.” NeurIPS. Empirical justification for Bayesian bandit-style exposure.

​
Code

  • Library: platform/src/lib/ml/maturity.ts
  • Tests: platform/src/lib/ml/__tests__/maturity.test.ts
  • Call site: platform/src/lib/pipeline-runner.ts (applyMaturityRamp)