Figure 3: The saturating law’s 24-week prediction hits, and the activity-mixture model’s 32-week prediction misses. Both were registered internally. (a) At
k=24 the observed
0.4296 lies
0.005 below the saturating prediction and
0.066 below the log-linear one, so both internally registered levels sit above the outcome. The two forms were fit on the three-window matched ladder before any 24-week evaluation and are scored across a cohort shift (both scorings in § 4.5 ). The curve plots the log-space
R2 of the weekly-aggregate ridge probe (closed-form, no training seed) against history length
k=1..24 weeks. Every point is scored on the single four-window matched cohort (
n=372,778 ) and comes from the M24 extract alone, so no training-extract mismatch enters the curve. On the test calendar it flattens (
+0.00009 /week at
k=22→24 ), but a single exponential underfits it, with fit RMSE
0.0103 and an asymptote of
0.423 , below the observed
0.4296 . A mixture of per-decile saturation rates predicts this underfit inside the window, and panel (b) rejects the mixture’s extrapolation past
k=24 . Inset: the more active the customers, the sooner the curve bends. The top three deciles, 10, 9 and 8, reach the half-saturation point
k1/2 of their saturating fits (circles) at
3.9 ,
5.2 and
10.0 weeks. In the fits, deciles 6–7 bend beyond the window (
k1/2 of
27.1 and
24.7 weeks), and the bottom five are still climbing at the window’s edge. This is a claim about curve shape, not level, and it rests on the fitted parameters in § 2.6 . Curves show per-decile refit
R2(k) ,
k=1..16 , from the M16 extract on the same matched cohort, shaded by prior-activity decile. Each (decile,
k ) pair has one ridge fit, evaluated on that decile’s test rows. All ten deciles are fitted, but only deciles
1 ,
5 ,
8 ,
9 and
10 are drawn, because the lower eight lie within
0.051 of one another at
k=16 and their strokes would overlap. (b) The 32-week miss, drawn against the falsification region of the internally registered extension. This panel uses a separate M32 extract, scored on the matched-cohort rows that also fall in the M32 test split (
94.2% coverage,
n=351,174 ). Its level differs from panel (a) by
−0.033 , an offset between extracts in which the M32 training data and the scored rows change together. The internally registered quantity is therefore the gain within this extract,
Δ(24→32) , anchored at the observed
k=24 point. The activity-mixture model predicted
Δ=+0.0142 [0.0126,0.0231] , the internally registered band and the only interval that panel (b) draws. Its lower endpoint is the calibrated and its upper endpoint the raw
90% bootstrap bound, so it is not a single
90% interval. The flat-tail null is
≤+0.001 . The observed gain is
Δ=−0.0004 , with 95% CI
[−0.0014,+0.0004] from the paired customer-clustered bootstrap on the same extract and rows. This observed interval reflects row-sampling uncertainty at the three test anchors, not variation across calendar anchors (§ 2.3 ). The curve enters the flat-tail null, and the observed interval excludes the entire internally registered range (discriminating threshold
0.007 ). The mechanism check fails in the same direction. The three lowest deciles show a mean within-decile gain of
0.018 , against a predicted
0.045 –
0.059 .