Research: Adaptive Page Length & Summary Systems

Problem Statement

The wiki has ~625 pages ranging from empty stubs (0 words) to comprehensive analyses (10,000+ words). Different readers want different levels of depth:

• Skimmers want a 2-sentence takeaway
• Decision-makers want a structured 200-word summary with key numbers
• Researchers want the full article with all evidence and nuance

Currently every reader gets the same page. The wiki has a two-tier summary system (description and llmSummary in frontmatter) and infoboxes, but these aren't doing enough to serve short-attention readers.

Current State

Page Length Distribution

Median: ~2,200 words. Mean: ~2,274 words. The longest page (bioweapons) is 10,646 words.

What Exists Today

Two-tier summaries:

• description (frontmatter): 1–2 sentences, ~100–200 chars. Used in meta tags and OG cards.
• llmSummary (frontmatter): AI-generated, 200–500 chars. Shown in PageStatus component, EntityLink hover tooltips, and InfoBox fallback.

InfoBox:

• Right-floating card (280px desktop, full-width mobile) showing entity metadata, key facts, related entries, ratings, links.
• Description is CSS-truncated to 3 lines (line-clamp-3) with no "read more" button.

PageStatus banner:

• Shows quality/importance badges, word count, and llmSummary text above the article.
• Currently a static block — no collapse/expand behavior.

Table of Contents:

• Sticky sidebar on desktop, auto-generated from H2/H3 headings.
• Scroll-tracking via IntersectionObserver.
• Only renders if 3+ headings exist.

Key gaps:

• No interactive expand/collapse on article sections.
• No user preference for "show me the short version."
• InfoBox description truncation has no expand affordance.
• Same rendering for 200-word and 10,000-word pages.

Approaches

Approach 1: Better Summaries & InfoBoxes (Low Effort, High Impact)

Don't change the page rendering model at all. Instead, invest in making the top-of-page experience much richer so skimmers never need to scroll.

What to do:

1. Structured llmSummary: Instead of a single paragraph, generate summaries with a consistent structure:

   • One-liner (what is this thing, in one sentence)
   • Key numbers (2–4 bullet points with the most important quantitative claims)
   • Bottom line (so-what for a decision-maker)
   • Store this as structured YAML in frontmatter rather than a single string.

2. Expand the InfoBox description: Add a "Show full summary" toggle to the InfoBox so the line-clamp-3 truncation is interactive. The llmSummary text is already there — just hidden by CSS.

3. Quick Assessment tables: Many KB risk pages already have a "Quick Assessment" table at the top. Standardize this across all page types (not just risks) and enforce it in page templates.

4. Key Takeaways component: A new MDX component (<KeyTakeaways>) that authors place after the first heading. Renders as a highlighted box with 3–5 bullet points. Unlike the auto-generated llmSummary, this is hand-curated.

Pros: No architectural changes. Works with static generation. Improves every page visit. Incremental — can improve one page at a time.

Cons: Doesn't actually shorten pages. Skimmers still see a long scrollbar. Every page still ships its full HTML.

Approach 2: Collapsible Sections (Medium Effort)

Keep full SSG but make sections below the overview collapsed by default on long pages.

What to do:

1. Auto-wrap H2 sections in a <CollapsibleSection> component during MDX compilation. The Radix UI Collapsible primitive already exists in the codebase (app/src/components/ui/collapsible.tsx).

2. Default behavior by page length:

   • < 1,500 words: all sections expanded (short page, no need to collapse)
   • 1,500–4,000 words: Overview expanded, rest collapsed
   • > 4,000 words: Overview expanded, rest collapsed, with a "Expand all" button

3. Remember user state: Use localStorage (pattern already established with DevModeToggle and InfoBoxVisibility) to remember which sections the user expanded.

4. Section word counts: Extend build-data.mjs to compute per-section word counts and display them as subtle badges on collapsed headers ("Risk Assessment — 840 words").

Pros: All content still pre-rendered (SEO-friendly). Readers see a manageable page. Existing Radix primitives reduce build effort. Section-level anchors and TOC continue to work.

Cons: All HTML still ships to the client (no payload savings). Requires an MDX plugin or remark plugin to auto-wrap sections. Testing the collapse/expand UX across 625 pages needs care.

Approach 3: View Mode Toggle (Medium-High Effort)

Give readers a "Summary / Standard / Full" toggle that controls how much of the page is shown.

What to do:

1. Three view modes:

   • Summary: InfoBox + structured summary + key takeaways only (~200–400 words). No article body.
   • Standard: Overview section + Quick Assessment + collapsed detail sections (~800–1,500 words visible).
   • Full: Everything expanded, as today.

2. A <ViewModeProvider> context (following the InfoBoxVisibilityProvider pattern) wrapping the page layout. A toggle in the page header or breadcrumb bar.

3. Persist preference in localStorage. Default to "Standard" for pages over 3,000 words, "Full" for shorter pages.

4. URL parameter support: ?view=summary so links can point to a specific view. Useful for sharing with non-technical audiences.

Pros: Directly addresses the "different readers want different depths" problem. URL parameters make it shareable. Could eventually be personalized per-user.

Cons: Significant UX design work. Need to ensure "Summary" mode is genuinely useful (not just a worse version of the page). Three modes means three things to test and maintain.

Approach 4: Per-User LLM Rewriting (High Effort, Experimental)

Dynamically rewrite pages using an LLM to match each reader's desired length and expertise level.

What to do:

1. On-demand summarization API: A serverless function that takes a page ID, target length (e.g., 200 words), and audience (e.g., "policymaker" / "ML researcher"), and returns a rewritten version using Claude.

2. Caching layer: Cache rewrites by (page_id, length, audience) tuple. Pre-generate common variants at build time for high-importance pages.

3. UI: A slider or dropdown: "Show me this in: 1 paragraph / 1 page / full article" with a loading indicator while the rewrite generates.

Pros: Most flexible. Each reader gets exactly what they want. Could support multiple languages, reading levels, or domain expertise.

Cons: Expensive (LLM call per page view, even with caching). Latency on first view. Risk of LLM introducing errors or losing nuance. Breaks the "wiki as source of truth" model — which version is canonical? Requires moving away from pure SSG. Hard to quality-control generated summaries at scale.

Approach 5: Tiered Content Authoring (High Effort, High Quality)

Author pages at multiple depths deliberately, as part of the content pipeline.

What to do:

1. Extend page templates to require three tiers of content:

   • Tier 1 — Card (~50 words): Title + one-line description + key metric. Already roughly exists as description field.
   • Tier 2 — Brief (~300 words): Structured summary with key numbers, main arguments, and bottom line. New frontmatter field or dedicated MDX section.
   • Tier 3 — Full: The existing article.

2. Authoring tooling: Extend crux content improve to generate Tier 2 briefs from Tier 3 content. Run this across all pages to bootstrap.

3. Render by context: Card view in search results and related-pages lists. Brief view as default for pages over 3,000 words (with "Read full article" link). Full view on click-through.

4. Quality control: Grade briefs separately from full articles. A bad brief on a good article is worse than no brief.

Pros: Highest quality — human-reviewed summaries at each tier. Each tier is a first-class piece of content. Cards improve search/browse experience even without changing page rendering.

Cons: Multiplies authoring work by ~2x. Need to keep tiers in sync when content changes. 625 pages is a lot of briefs to write (though LLM-assisted).

Recommendations

The approaches aren't mutually exclusive. Here's a phased plan:

Phase 1 — Improve what exists (Approach 1)

This is the highest-leverage, lowest-risk starting point:

• Restructure llmSummary into a structured format (one-liner, key numbers, bottom line). Run the crux authoring pipeline to regenerate summaries for all pages.
• Add an expand toggle to the InfoBox description (currently just CSS line-clamp-3).
• Standardize Quick Assessment tables across all page types (risks already have them).
• Create a <KeyTakeaways> MDX component for hand-curated bullet summaries.

Phase 2 — Collapsible sections (Approach 2)

Once summaries are solid, make long pages less intimidating:

• Build a remark plugin that wraps H2 sections in <CollapsibleSection>.
• Default-collapse sections below Overview on pages over ~2,000 words.
• Add section word counts to collapsed headers.
• Store expansion state in localStorage.

Phase 3 — View mode toggle (Approach 3, partial)

Add a Summary/Full toggle (skip the "Standard" middle mode to reduce complexity):

• Summary mode shows structured summary + InfoBox + Key Takeaways only.
• Full mode shows the complete page (with collapsible sections from Phase 2).
• Persist preference. Support ?view=summary URL parameter.

Phase 4 (optional) — Tiered authoring (Approach 5)

If Phase 1–3 show strong user demand for short content:

• Extend page templates to include a Tier 2 brief section.
• Use crux content improve to auto-generate briefs, then human-review.
• Render briefs as the default view for high-importance, high-length pages.

Skip for now: Approach 4 (per-user LLM rewriting). The cost, latency, and quality-control challenges outweigh the benefits given the wiki's static architecture. If the wiki moves to a server-rendered model for other reasons, revisit.

Technical Feasibility Notes

Existing infrastructure that helps:

• Radix UI Collapsible and Tabs components are already in the codebase
• InfoBoxVisibilityProvider context establishes the toggle-with-localStorage pattern
• DevModeToggle proves localStorage preferences work across sessions
• extractHeadings() in mdx.ts already parses H2/H3 structure at compile time
• build-data.mjs already computes page-level metrics; extending to section-level is straightforward
• crux content improve can be extended to generate structured summaries

What would need to be built:

• Phase 1: Structured summary YAML schema, <KeyTakeaways> component, InfoBox expand toggle (~200 lines)
• Phase 2: Remark plugin for section wrapping, <CollapsibleSection> component, section metadata in build pipeline (~500 lines)
• Phase 3: <ViewModeProvider> context, header toggle, URL param handling (~300 lines)

Phase 1 Implementation Status

Phase 1 has been implemented. Here's what was built:

1. Structured Summary Schema

New structuredSummary frontmatter field with three sub-fields:

structuredSummary:
  oneLiner: "One sentence explaining what this thing is"
  keyPoints:
    - "First key quantitative or qualitative finding"
    - "Second key point"
    - "Third key point"
  bottomLine: "So-what for a decision-maker"

Files changed:

• crux/lib/rules/frontmatter-schema.ts — Zod validation schema
• app/src/data/index.ts — TypeScript Page interface
• app/scripts/build-data.mjs — Build pipeline passthrough

2. InfoBox Expand/Collapse Toggle

The InfoBox description was previously CSS-truncated to 3 lines (line-clamp-3) with no way to see the full text. Now it's interactive: if the text overflows 3 lines, a "Show more" / "Show less" button appears.

New file: app/src/components/wiki/InfoBoxDescription.tsx — Client component using useRef + useEffect to detect overflow, toggling line-clamp-3 on click.

3. KeyTakeaways MDX Component

A new <KeyTakeaways> component for hand-curated bullet summaries at the top of articles. Renders as a visually distinct card with an indigo accent (lightbulb icon + "Key Takeaways" header).

New file: app/src/components/wiki/KeyTakeaways.tsx Registered in: app/src/components/mdx-components.tsx

4. Structured Summary Rendering in PageStatus

When a structuredSummary is available, PageStatus renders it as:

• One-liner in bold at the top
• Key points as a bullet list
• Bottom line in a highlighted indigo box with a "BOTTOM LINE" label

Falls back to the existing flat llmSummary paragraph when no structured summary exists.

5. Example Pages

Structured summaries and KeyTakeaways added to:

• content/docs/knowledge-base/risks/scheming.mdx — Risk page with both structuredSummary and <KeyTakeaways>
• content/docs/knowledge-base/responses/interpretability.mdx — Response page with both

Next Steps for Phase 1 Rollout

• Run crux content improve across all pages to auto-generate structuredSummary fields
• Hand-review generated summaries for high-importance pages (importance >= 70)
• Add <KeyTakeaways> to the top 30 most-visited pages
• Consider making structuredSummary a required field in page templates for new pages