Validation against Bishop’s reference results

Every analytical output this library produces is checked, number by number, against Dr. Thomas A. Bishop’s VAS results (Minitab) for the same data. Not “inspired by” and not spot-checked — the chart centers, control limits, capability indices and loss-function components are compared to the reference and the run fails if any of them disagree.

1547 assertions pass, 0 fail, 4 have no reference value yet.

Generated by validation/e2e_bishop_report.py from processbehavior 0.4.0.

Coverage¶

Run	Data	Analytic design state	Assertions	No reference yet	Result
`PM SDS 1`	`PBTESTDATABASE_T100.csv`	ADS 1	105	0	✅ all pass
`PM SDS 2`	`PBTESTDATABASE_T100.csv`	ADS 2	105	0	✅ all pass
`PM SDS 3`	`PBTESTDATABASE_T100.csv`	ADS 3	105	0	✅ all pass
`PM SDS 4`	`PBTESTDATABASE_T100.csv`	ADS 1	105	0	✅ all pass
`PM SDS 5`	`PBTESTDATABASE_T100.csv`	ADS 2	105	0	✅ all pass
`PM SDS 6`	`PBTESTDATABASE_T100.csv`	ADS 3	105	0	✅ all pass
`PM INERT`	`PBTESTDATABASE_T100.csv`	ADS 1	105	0	✅ all pass
`PM INERT SDS 2`	`PBTESTDATABASE_T100.csv`	ADS 2	105	0	✅ all pass
`PM INERT SDS 3`	`PBTESTDATABASE_T100.csv`	ADS 3	105	0	✅ all pass
`PM INERT SDS 4`	`PBTESTDATABASE_T100.csv`	ADS 1	105	0	✅ all pass
`PM INERT SDS 5`	`PBTESTDATABASE_T100.csv`	ADS 2	105	0	✅ all pass
`PM INERT SDS 6`	`PBTESTDATABASE_T100.csv`	ADS 3	105	0	✅ all pass
`PM SDS 1 KNOWN`	`PBTESTKNOWNEFFECTS_T100.csv`	ADS 1	101	4	✅ all pass
`Medicare ACO (PCE)`	`aco_per_capita_expenditure.csv`	ADS 2	186	0	✅ all pass

PM SDS 1-6 are one fill-weight dataset with deletions; SDS 4-6 are incomplete grids, analysed as the complete design that survives tidying (ADS 1-3). PM INERT is pure noise with the same six missing-data patterns. The known-effects data are built from known condition, time and interaction effects plus noise. The Medicare data are per-capita expenditure for 24 organisations over 4 years, one reading each (issue #114).

What is covered¶

Charts — center line and both natural process limits, for the primary chart and for each valid chart x residual pair at that design state. Where subgroup sizes differ, VAS prints “UNEQUAL” instead of a limit; there the limits of every subgroup are read from the limit lines VAS draws, to within 1.5 drawing units (about 0.01-0.1, by chart), and each line counts as one assertion.
Capability — Pp, Ppk (upper and lower), Cp, Cpk, and the percentage of the distribution beyond each specification limit.
Loss function — the Taguchi loss decomposition and its components.

Values with no reference yet¶

The limits of the two factor S charts (pages 25 and 27) in the known-effects run, whose subgroups have 1000 and 2000 values. Its VAS run of 6 October 2026 held α_N (10-12 manual Eq 10-11) at 0.9995 from N = 343 up, where Minitab’s GAMMA function overflows, so its limits follow N = 500 instead of the subgroup’s own N. Dr. Bishop’s reruns of the PM runs (13 October 2026) compute α_N for every N, and every R6 S chart in them is checked and matches; the known-effects run waits for the same rerun.

Reproducing this¶

python validation/e2e_bishop_report.py

That writes this page and a detailed HTML report (validation/e2e_bishop_report.html) with every compared value, expected beside actual, grouped by run.