Every card is compared at respondent level: shares pooled within each group, a Wilson interval on each share, and a Newcombe interval on the difference. Each market’s own value is kept beside every aggregate, and every contrast is recomputed with each market dropped in turn — a gap that reverses when one market is removed rests on that market, not on the group.

280 contrasts are computed across the page, so some would clear an interval test by chance alone. Before a card may claim a difference, its contrast is screened with a Benjamini–Hochberg false-discovery control at 5%, and the gap must reach at least 10 points. Individual test statistics are not reported: MAD publishes effects, intervals and ranges rather than p-values.

A card claims the groups are the same only when its interval fits inside ±15 points. Anything wider is recorded as not resolved, which is a statement about the sample rather than about the world.

With two markets in Group A against four in Group B there are only fifteen ways to split six markets, so a market-level rank test bottoms out at an exact two-sided probability of 0.133. No market-level result can reach conventional significance under this design, however large the effect. Complete separation of the markets is therefore reported as corroboration for a respondent-level contrast and never as a finding on its own.

The composition adjustment reweights both groups to the pooled respondent profile — stratified by organisational discipline for Organisational Adoption and academic programme type for Education Landscape. It is derived at build time from the sanitised per-respondent files, because the published data holds only per-option totals and no cross-tab between two questions. It is withheld where fewer than 60% of respondents sit in profiles shared by both groups.

These are descriptive contrasts between two groups of markets, not causal evidence. The groups differ in more than their policy environment, and nothing here isolates policy as the cause of any gap.