Insights

Question 1

How close are issues within and across trees?

Issues are closest within the same subcategory. Average closeness is 25.8 within subcategories, 16.3 across branches of the same tree, and 3.9 across different trees, on a 0–100 scale. Strong cross-tree matches often reflect repeated concepts, narrower scopes, or a problem and its safeguard. These links are useful starting points for taxonomy review and further research.

There are 154 cross-tree pairs scoring at least 50. A low average between whole trees can still contain a strong individual connection. The cutoffs 25 and 50 are screening values, not evidence grades.

Taxonomy density matters: 446 of the 1,139 issue labels sit in the Defamation, Racial Hierarchy and Racial Hate trees, while Disability has three and Education four. Density affects rankings; it does not measure social importance. The workbook and detailed-data download contain all subcategory results and complete matrices.

Question 2

How close are intersectional issue groups to one another?

Most retained higher-order website groups stay within a single subcategory: 2,127 of 2,492 K = 3–7 groups, or 85.4%. Their median internal cohesion declines from 30.1 at K = 3 to 22.6 at K = 7. Similar groups often share force-related or records-and-sourcing concepts. The outside-in search adds 5,153 unique candidates for further investigation.

These are differently selected K-level pools, not a time trend or an experiment proving that adding an issue causes a particular score change.

“Within one subcategory” and “cross-tree” are not exhaustive categories: groups can span subcategories of one tree. Only 70 of 340 retained K = 7 groups have every constituent pair scoring at least 25.

The candidate universe is 16,897 retained website groups + 5,099 new greedy cross-tree groups + 54 new unique source-informed theme groups = 22,050 unique groups. The ten themes contain 60 K-level entries; some reuse existing candidate sets, so unique groups are counted separately from theme entries.

Selected same-K, non-overlapping comparisons

Each example comes from exported nearest-neighbor results. It is not a claim that an unrestricted global maximum was identified for every K.

The study computed 243,090,225 unordered group comparisons across 22,050 candidates. The downloadable package contains summaries, nearest-neighbor outputs, the issue kernel and a query script, not a flat 243-million-row export. Higher-order discovery is bounded; this is not an exhaustive evaluation of every possible K = 3–7 combination.

Question 3

Which stakeholder groups are closest to these intersections?

Stakeholder relevance becomes more useful when the affected population and institutional role are specified. This study compares the website’s 287-cohort roster with 38 additional outside-in research profiles. It separates people directly exposed to a process from possible support roles and decision or resource roles. Useful leads include paid care workers for migration–care, housing-burdened households for housing exclusion, and women journalists for publication and digital safety.

A circle is an operational role, not an actual social network or a claim about influence.

The website roster contains 287 profiles: 71 role-coded profiles are scored and 216 geography-only profiles remain unscored. Those 216 comprise 194 dual-national country cohorts, 17 metropolitan lenses and five state cohorts. Nationality or place alone does not establish relevance to a particular issue mechanism.

8,707 / 22,050 groups have a positive website direct-circle candidate; 9,406 / 22,050 have an outside-in direct-circle candidate. These are model coverage counts, not evidence that other groups have no relevant stakeholders. Website and outside-in profiles remain separate; the 38 outside-in profiles are not additions to the canonical roster or governed relationship graph.

K = 7 direct-circle overview

Theme and K inspector

Select one of the 60 supplied theme/K records. Internal cohesion, weakest pair, connection bottleneck and distinct trees use separate labels. The connection bottleneck is the weakest edge of a maximum spanning tree and can be positive even when a nonadjacent pair scores zero.

Coverage rules: A null score is shown as “Not established”; numeric zero remains a model result. “Tied among N candidates” means stable ID ordering selected the displayed candidate. “Descriptor coverage known < K” flags incomplete inputs. Raw feature overlap does not validate a match or override a null score.

The site separately publishes 71 governed relationships covering 48 issues and 17 cohorts. That evidence layer is not the same as the 71 role-coded profiles. Only 129 candidate groups have every constituent covered for the same cohort, and even that is constituent coverage rather than proof of a joint K-way effect.

Methods and limitations

How the screening model works

Outside-in means an independent comparison of issue descriptions and institutional features, without a bonus for placement in the same website branch.

Downloads and sources

Read, inspect or reproduce the analysis

The detailed-data ZIP is the larger download. Files are served as ordinary same-origin assets and work without JavaScript. Exact metadata is available in the download manifest.