Importance Ranking System

Page importance scores (0-100) are derived from ordered rankings — lists of all pages sorted by importance. The ranking is the source of truth; numeric scores are computed from position.

Two Dimensions

The system maintains two separate rankings that measure different things:

These are intentionally different. A broad overview page like "AI Alignment" ranks high for readership but low for research (it compiles existing knowledge rather than generating new insights). A narrow page like "Mesa-Optimization" or "Sleeper Agents" might rank higher for research because deeper investigation could reveal critical findings.

Why Rankings Instead of Direct Scores?

The previous system asked "rate this page's importance from 0 to 100" — but arbitrary scores are hard to calibrate. Is a 72 really more important than a 68? What does a 50 mean in absolute terms?

Rankings solve this by forcing relative comparisons: is page A more important than page B? This is a much easier judgment call, and it ensures every score is meaningful relative to every other score.

The 0-100 scores are then derived mechanically: position 1 maps to ~95, the last position maps to ~5, with linear interpolation between.

How It Works

Batch Sort + Merge

For ranking all ~645 pages:

1. Phase 1 — Batch sort: Split pages into groups of ~25. For each group, a single LLM prompt sorts them by importance. (~26 prompts)
2. Phase 2 — Merge: Insert pages from remaining batches into the growing master ranking using binary search (each insertion takes ~10 pairwise comparisons). (~5,000 comparison calls)
3. Phase 3 — Verification: Slide a window of 20 pages across the ranking, re-sorting each window to fix local inversions from merge noise. (~63 prompts)

Cost

A full ranking run costs approximately $1-2 using Haiku. Breakdown:

• 26 batch sort prompts (≈$0.13)
• ≈5,000 binary search comparisons (≈$1.20)
• ≈63 verification windows (≈$0.30)

Prompts

The readership and research dimensions use different system prompts that change what the LLM optimizes for:

• Readership: Centrality, foundational dependency, real-world relevance, breadth
• Research: Insight potential, neglectedness, decision relevance, crux resolution

CLI Commands

# View rankings
pnpm crux importance show --top=30                        # Readership (default)
pnpm crux importance show --dimension=research --top=30   # Research

# Rerank all pages (takes ~40 min, costs ~$1-2)
pnpm crux importance rerank --all --apply
pnpm crux importance rerank --dimension=research --all --apply

# Test with a sample first
pnpm crux importance rerank --sample=20
pnpm crux importance rerank --dimension=research --sample=20

# Fix local inversions in existing ranking
pnpm crux importance rerank --verify --apply

# Write derived scores to page frontmatter
pnpm crux importance sync --apply

# Bootstrap from existing importance scores (one-time setup)
pnpm crux importance seed --apply

Relationship to Other Systems

• Critical Insights: The research ranking aligns with the Critical Insights framework's emphasis on surprising, important, and neglected topics. Pages ranking high in research importance are natural candidates for insight extraction.
• Gap Analysis: The gap system (pnpm crux gaps) uses importance scores to prioritize which pages need more insights. Updated importance scores feed directly into gap prioritization.
• Update Schedule: The update scheduler uses importance to determine which stale pages to refresh first.
• Search Ranking: Importance provides a tiebreaker boost in search results.
• Homepage: Featured content on the homepage weighs importance 2x.

Files