How the map is organized

Theme definitions & evaluation plan · November 7, 2026 · Prepared with Codex assistance

Five provisional themes help readers find related questions. Independent human evaluation of the assignments remains open.

node&norm assigned the themes with Codex assistance, using the records’ existing labels, summaries and research questions. The published assignment record preserves the definitions, primary and secondary themes, and reasons for placement.

Open the connections map →

The five themes

ThemeQuestion it helps readers explore
Authority & delegationWho or what may act, within which limits, and how permission can be delegated, constrained or withdrawn.
Human judgment & oversightThe information, expertise, working conditions and opportunities people need to understand, question and influence an AI-assisted decision.
Evaluation & verificationMethods for examining capabilities, risks and controls under stated test conditions, with findings and uncertainty kept in scope.
Evidence & traceabilitySources, reasons, records and versions that make decisions and research claims inspectable and reveal what remains unknown.
Responsibility & repairAnswerability for consequences, routes to challenge, responses to concerns and evidence that corrections reach affected people and decisions.

How placement works

Each record has one primary theme for placement. Secondary themes identify other reading contexts. The concepts can overlap: withdrawing permission may also raise questions about evidence or repair. Secondary themes create no additional graph connections; each drawn link retains its own recorded meaning.

The five themes appear as labeled territories. Research topics sit in a central ring and retain their primary theme in the record card and theme counts. Shape and tone identify record type: filled circles for research, diamonds for governance, triangles for risks and rings for incidents. Position and proximity help navigation and do not measure risk, similarity, evidence strength or research maturity. Theme assignments leave source claims, legal status and existing relationships unchanged.

In the expanded trace, columns group records by type. Reviewed direct links are distinguished from “one step on”: an outward continuation through an explicitly connected intermediate record. This is a selected reading route, not an exhaustive two-step neighborhood, inferred direct relationship, causal claim or validation. The record card provides equivalent lists with named intermediates on all screen sizes.

What we plan to test

The proposed evaluation asks whether people can apply the definitions consistently and use the map without misreading its visual cues. Recruitment, independent ratings and visitor testing have not been completed under this plan.

01

Define the boundaries

Prepare a versioned coding guide with inclusion and exclusion rules, examples, overlapping cases, and options for no fit or insufficient information. Pilot it on records spanning the five themes and the research, governance, risk and incident groups.

02

Compare independent judgments

Ask human reviewers outside the assignment team to rate the same source materials independently, with current labels, colors and rationales hidden. Record a preferred primary theme, other fitting themes, uncertainty and reasons. Preserve disagreements before any discussion.

03

Test a fresh sample

Freeze the revised guide and evaluate records held out from the pilot, including ambiguous cases. Report agreement, coverage, confusion between themes and no-fit rates by theme and record type, with uncertainty. Compare existing assignments only after independent ratings are fixed.

04

Check visitor understanding

Ask visitors to find a record, follow a connection and inspect its source on desktop, mobile and with keyboard navigation. Check their understanding of colors, nearby dots and link meanings. Record task completion, errors and mistaken causal or risk-ranking interpretations; revise and retest with new participants.

Criteria before any validation claim

Before the fresh evaluation, publish the sample size and selection rules, reviewer eligibility, agreement measure, uncertainty method, minimum acceptable agreement and coverage, and rules for resolving or retaining ambiguity. Numerical thresholds remain to be specified and justified in that protocol. They must be fixed before the evaluation results are seen.

A claim of validation would require meeting those criteria, publishing the results and limitations, and independent peer review of the method and its interpretation. Revisions after a failed criterion would need a new untouched evaluation sample. Peer review and agreement alone would not establish an objective taxonomy or remove editorial judgment.

Any resulting claim must identify the tested version, collection, readers and navigation task. Research validity, legal applicability and effectiveness in deployed institutions require their own evidence.

Status: proposed evaluation, not a completed validation study. Existing assignments and research findings are unchanged. Future changes to definitions or placements should preserve a versioned explanation and the earlier record.