Every number on this page is computed by the Chapter 4 notebook's own code. The adoption history of step 3 is the notebook's bench, drawn with the lab's random seed 26104, not a Meridian record. Your notes stay on this page until you leave it; copy the memo before you go.
The board is about to approve Meridian's digital program on the strength of a planning model that fits seven years of history well. A director asks the question no one has answered: how far can that model be trusted for a program that will take Meridian somewhere it has never been? You are on the planning team.
Use AXIOM to test the planning model the way the chapter recommends: check its chain, test its roll-ups, map where its validation applies, and set rules for when it must be revisited.
Worked Example 4.1: Industrial Systems' engineers in two pools. Each year 5 percent of legacy engineers leave and 3 percent are retrained into software; 4 percent of software engineers leave. Unit A has 400 legacy and 100 software engineers, Unit B the reverse.
In the model-chain builder, follow the engineering-headcount question of Worked Example 4.1 from diagram to running code, and note what each verification checkpoint tests.
Table 4.1's four model types; Definition 4.1; Table 4.4's verification items.
Each link translates its predecessor into a stricter medium (Table 4.1), and each verification item tests one translation (Table 4.4): that the mathematics renders the logic (the pools keep 92 and 96 percent, so the declared shares leave), that the code implements the mathematics (the recursion against its closed form), and that the model never leaves its own admissible region (no pool turns negative). Verification asks whether the model is built right; validation, in steps 3 and 4, whether it is the right model.
Table 4.1; Figure 4.1; Definition 4.1; Table 4.4; Appendix 4.A; notebook cell 4.
In the aggregation tester, roll Industrial Systems' engineers up to a single total and watch whether the summary model and the detailed model agree. Add the skill mix to the summary and test again.
Theorem 4.2's linear test: a one-row summary p is a state of its own exactly when the dynamics map every configuration it cannot see (p·v = 0) to one it still cannot see (p·A·v = 0). The total-only model, calibrated on Unit A's first year, multiplies the total by 476/500 = 0.952 a year.
On the model bench, fit a simple and a rich model to the same adoption history and slide the rich model's complexity up. Compare the fit on the history with the error on the later quarters.
The bench: adoption of a digital product (percent of target users) following an S-curve with ceiling 60 percent, plus noise of 1.5 points, quarters 1–24. Both models are fitted by least squares to quarters 1–16 and judged on quarters 17–24. The book prints no such history; these are the notebook's labelled settings.
In the validation-domain explorer, fit two models to the same Meridian record, Unit A's first year, confirm that they fit equally well, then run both at skill mixes the record never saw and watch them separate outside the shaded region.
Next year's total for a unit of 500 engineers. The total-only model says 0.952 × 500 whatever the mix; the detailed model, H × (0.95 + 0.01 s), depends on the software share s. The shaded band is the record: the mix Unit A's history visited, 20 to 33 percent (Definition 4.6; Figure 4.3).
In the lifecycle monitor, replay five years of use with drift, and set the triggers that should send the model back for revalidation.
The five-year plan is the total-only model; what happens is the detailed model, for each unit. The indicators are the plan's error and the software share's distance from the 20 percent mix the model was calibrated on. The book sets no tolerances; these are settings to vary.
A drift trigger on the input mix fires early for Unit A, while its plan error is still small: it warns before the error does. For Unit B the mix is outside the record from the start, so the model should never have been used there without revalidation. The book adds an event trigger: a transformation itself, the program's start or end, sends the model back at once (Section 4.6; Figure 4.4's seven stages, from specify to revise or retire).
Section 4.6; Figure 4.4; Table 4.4's monitoring item; notebook cell 18.
The total-only summary drifts while the mix-aware summary holds (Theorem 4.2); the better fit on history can be the worse forecast; and two models indistinguishable on the record can disagree sharply beyond it (Theorem 4.3).
The readings and units, the assumption classes of Table 4.3, the validated domain, and the revalidation triggers.