Website Consistency Audit (February 2026)

This report documents a systematic audit of the wiki's consistency, quality distribution, navigation structure, and cross-linking completeness. It covers 672 non-internal MDX pages and identifies specific problems with prioritized recommendations.

Executive Summary

The wiki has strong foundational infrastructure (templates, grading engine, citation tracking, data-driven navigation) but suffers from inconsistent adoption of its own systems. The core problems are:

1. 15% of scored pages are stubs or near-stubs, many covering the wiki's most important topics
2. Page structure varies wildly within the same entity type — section names, table formats, and component usage differ on every page
3. Frontmatter fields are inconsistently populated — critical fields like numericId (16%), subcategory (71%), and clusters (80%) have major gaps
4. Navigation has dead zones — 20+ pages have no sidebar and 89 KB pages can't appear on the homepage
5. Two citation systems coexist — <R id="..."> and [^N] footnotes are used interchangeably with no clear rule

The wiki is in a state where the systems are well-designed but the content was created across many sessions with insufficient enforcement of those systems.

1. Quality Distribution

Quality Score Breakdown (622 pages with scores)

81 additional pages have no quality field at all, including overview/hub pages with importance scores above 90.

Worst Quality/Importance Mismatches

These pages have the largest gap between how important they are and how little content they have:

is-ai-xrisk-real has the highest importance score in the entire wiki (93) but only quality 12. This is arguably the single most important page on an AI safety wiki and it barely exists.

60 total pages have quality under 30 with importance over 50.

2. Structural Consistency by Page Type

Risk Pages (74 pages)

What's consistent: All have an Overview section, all use EntityLink, all have DataInfoBox and DataExternalLinks.

What's inconsistent:

• Section names vary on every page. The template requires "How It Works" — actual pages use "How AI Could Help Attackers", "Technical Capabilities & Development", "How Racing Dynamics Work", "The Autonomy Spectrum and Human Control", etc.
• Risk Assessment table columns differ. Every sampled page has different headers: Risk Factor | Assessment | Evidence | Timeline vs. Dimension | Assessment | Evidence vs. Factor | Severity | Likelihood | Timeline | Trend vs. Dimension | Rating | Justification.
• "Responses" section uses 4 different names across 5 pages.
• "Key Uncertainties" uses 4 different names.
• Only 1 of 5 sampled pages has a "Quick Assessment" table — this is an emerging pattern that has not been rolled out.

Response/Approach Pages (162 pages)

What's consistent: "Quick Assessment" table appears on most pages. EntityLink is universal.

What's inconsistent:

• Quick Assessment table columns vary: Dimension | Assessment | Evidence vs. Aspect | Status vs. Dimension | Rating | Notes vs. Aspect | Details.
• "How It Works" is required by template but uses 4 different label variants.
• "Risks Addressed" (required) appears on only 3 of 5 sampled pages — policy pages skip it.
• Citation style split: Some pages use <R id="...">, others use [^N] footnotes, with no clear rule.
• DataInfoBox presence varies — only 2 of 5 sampled pages include it.

Organization Pages (121 pages)

Widest variation of any type. The organizations directory contains a mix of:

• Proper organization profiles (deepmind, miri, openai)
• Data table pages (entityType: table)
• Concept analyses (entityType: concept) about organizational topics
• Stub redirects
• Overview aggregation pages (entityType: overview)

Among proper organization pages, section count varies from 8 to 20 H2 sections. Section naming is inconsistent ("Historical Evolution" vs. "History" vs. "Organizational Evolution"). Quick Assessment appears on only 1 of 3 sampled proper org pages.

Person Pages (45 pages)

• 2 of 5 sampled pages lack an "Overview" section, using "Background" instead
• Risk Assessment tables appear on 3 of 5 (not a template requirement)
• Import sets and component usage vary across every page

Model Pages (89 pages)

• The highest-quality model pages (scaling-laws quality 92, ai-timelines quality 95) have the sparsest frontmatter — they predate the standardized template system
• DataInfoBox presence varies — some include it, some omit it entirely
• All sampled model pages use [^N] footnotes rather than <R> references

3. Frontmatter Consistency

Field Coverage

Specific Issues

Rating dimensions are inconsistent. The standard 4 dimensions (novelty, rigor, actionability, completeness) are on most rated pages. But 122 pages (21%) have extra dimensions (concreteness, focus, objectivity) introduced around December 2025 and not rolled out to other pages.

Legacy/duplicate field names exist:

• 16 pages use importance instead of readerImportance/researchImportance
• 1 page uses lastUpdated instead of lastEdited
• 2 pages use todo (singular) instead of todos (plural)
• 5 pages have a non-standard entityId field

entityType vs. directory mismatches: 29 pages, including 6 genuine mismatches where entityType: concept or entityType: table pages live in the organizations/ directory.

Subcategory quoting is inconsistent: 9 pages use quoted values ("governance-legislation") while all others use unquoted. Functionally identical in YAML but a formatting inconsistency.

8 singleton subcategory values exist (used by only 1 page each), suggesting incomplete classification.

4. Navigation and Cross-Linking

Sidebar Coverage Gaps

~20 pages have no sidebar navigation. These are reachable only via direct URL or cross-links.

Additionally, 3 pages at the KB root level (architecture-scenarios-table.mdx, deployment-architectures-table.mdx, directory.mdx) don't belong to any section sidebar.

Homepage Coverage Gaps

The homepage features 4 topic clusters (AI Safety, Governance, Biorisks, Epistemics) that spotlight pages. Key gaps:

• Biorisks cluster has only 4 pages assigned (vs. 116 for AI Safety)
• 89 KB pages (16%) have no cluster assignment and can never appear on the homepage
• No direct links to KB section index pages (Risks, Responses, Organizations, People, Models, etc.)

EntityLink Coverage

• 81% of all MDX files use <EntityLink> — strong overall
• 93% of KB content pages have cross-links
• 40 KB content pages (7%) have zero EntityLinks — concentrated in organizations, intelligence paradigms, and stub pages

Related Entries Coverage

66% of organizations and 100% of epistemic-orgs-projects entities have zero explicit related entries. The automatic relatedGraph compensates somewhat via tag and similarity matching, but the signal is weaker.

Index Page Consistency

14 of 15 KB section index pages follow a consistent template (Overview, Categories, EntityLinks, Why This Matters). The incidents index is an outlier — only 28 lines, no EntityLinks, no content listings.

5. Citation and Content Integrity

Two Citation Systems

The wiki uses two incompatible citation approaches:

1. <R id="..."> component references — structured, linked to the resources YAML layer
2. [^N] markdown footnotes — inline, not connected to any data layer

Roughly half the pages use each system, with some pages mixing both. Model pages and policy pages tend toward footnotes; risk and technical response pages tend toward <R>. There is no documented rule for when to use which.

Content Confidence System

The three-tier hallucination risk banner is well-designed. However:

• Zero human reviews have been recorded (review tracking infrastructure exists but is empty)
• 475 total citations across the wiki; only 279 (59%) checked
• 6% error rate among checked citations (17 inaccurate + 13 unsupported)
• External links on at least one high-importance page contain hallucinated URLs

6. UX Issues

1. No table of contents on long articles (some are thousands of words with many sections)
2. No pagination on the Explore page (625+ cards load at once)
3. Search button is small and keyboard-shortcut-focused (Cmd+K); not immediately discoverable
4. Homepage doesn't disclose that content is AI-generated (individual pages have confidence banners but the homepage does not)
5. Browse MDX pages (/browse/) appear to be a legacy layer superseded by the /wiki app route — potential confusion

Prioritized Recommendations

Critical (Highest Impact)

1. Fill the top 20 quality/importance gaps. Pages like is-ai-xrisk-real (importance 93, quality 12), existential-catastrophe (importance 92, quality 0), and transformative-ai (importance 92, quality 0) are the wiki's biggest credibility risks. Use pnpm crux content create or pnpm crux content improve for each.

2. Standardize section names within each entity type. The template system defines alternateLabels but actual pages use labels not in the list. Options:

   • Enforce template section names via a new validation rule
   • Expand alternateLabels to cover all observed variants
   • Run a batch rename across existing pages

3. Standardize table column formats. Risk Assessment, Quick Assessment, and Key Links tables should use identical column headers across all pages of the same type.

High Priority

4. Populate missing numericId values. Currently 16.4% coverage. Run node apps/web/scripts/assign-ids.mjs to assign IDs to all entities/pages. This unblocks reliable <EntityLink> resolution.

5. Complete clusters tagging for the 89 KB pages without assignments, especially organizations (26% untagged) and responses (17% untagged).

6. Pick one citation system and migrate. Recommendation: standardize on <R> references since they connect to the resources data layer. Create a validation rule that flags [^N] footnotes.

Medium Priority

8. Add sidebar navigation to browse, guides, project, and insight-hunting sections. These 20 pages are currently navigational dead ends.

9. Fix entityType/directory mismatches. Move 6 misplaced pages to their correct directories or correct their entityType.

10. Clean up legacy frontmatter fields. Rename importance to readerImportance (16 pages), lastUpdated to lastEdited (1 page), todo to todos (2 pages). Remove non-standard entityId fields (5 pages).

11. Populate quality on the 81 pages missing it, especially overview/hub pages with high importance scores.

12. Standardize rating dimensions. Either roll out the extra dimensions (concreteness, focus, objectivity) to all pages or remove them from the 122 pages that have them.

Lower Priority

13. Strengthen organization cross-linking. 66% of organizations have no explicit relatedEntries. Add at least 3-5 related entries per entity.

14. Add table of contents to long articles (word count > 1500).

15. Merge or redirect the legacy /browse/ pages to /wiki equivalents.

16. Fix the incidents index page — add EntityLinks and content listings.

17. Review singleton subcategory values (8 values used by only 1 page each) for consolidation.