Delegation Risk

A structured approach to managing risk in delegation relationships—with applications to AI safety.

Read Online | Download PDF | Download EPUB

Overview

Every delegation involves risk. When you delegate a task—to an employee, a contractor, a software system, or an AI agent—you're accepting potential downside in exchange for capability.

This framework proposes structural constraints as the foundation for managing delegation risk:

• Delegation Risk = Σ P(harm) × Damage — quantify what you're betting on each delegate
• Decomposition — many limited components instead of one powerful delegate
• Least-X Principles — minimize capability, context, autonomy, connectivity
• Cross-domain methods — adapt proven approaches from nuclear safety, finance, and mechanism design

Documentation Structure

Section	Description
Getting Started	Introduction, core concepts, FAQ, glossary
The Framework	Delegation risk theory, accounting, decomposition
Applying the Framework	Principles, architecture, worked examples, tools
Cross-Domain Methods	Finance, nuclear safety, mechanism design
Case Studies	AI systems, human systems, Anomaly Chronicles
Deep Dives	Theoretical foundations, research background
Reference	Bibliography, protocols, roadmap

Stats

• ~100,000 words of content
• 99 documentation pages
• 146 diagrams

Development

This is an Astro site using Starlight.

# Install dependencies
pnpm install

# Start dev server
pnpm dev

# Build for production
pnpm build

Generating PDF/EPUB

The documentation can be exported as a PDF or EPUB book.

Prerequisites

# macOS
brew install pandoc
brew install --cask basictex  # or mactex for full install
npm install -g @mermaid-js/mermaid-cli

Generate

python3 scripts/build-pdf.py

This generates:

• delegation-risk-framework-book.pdf (~9 MB)
• delegation-risk-framework-book.epub (~9 MB)

Features:

• Proper sidebar ordering matching the website
• Pre-rendered Mermaid diagrams
• Internal links converted to PDF anchors
• External links as footnotes (print-friendly)
• Colored headers, running page headers
• Custom EPUB styling

Versioning

The documentation uses an auto-incrementing version system that keeps PDF/EPUB and website in sync.

Version format: 1.0.3+abc1234

• 1.0 — Major.minor version (manually controlled)
• .3 — Patch version (auto-increments when content changes)
• +abc1234 — Git commit hash (for traceability)

How it works

1. When you run python3 scripts/build-pdf.py, it:

   • Hashes all documentation content
   • Compares with the stored hash in version.json
   • If content changed → auto-increments patch version
   • Stores the git commit hash for traceability
   • Embeds the version in PDF/EPUB title page

2. The website reads version.json and displays the version in the footer

3. Both PDF/EPUB and website show the same version

Manual version bumps

For significant changes, manually edit version.json:

{
  "major": 1,
  "minor": 1,
  "patch": 0,
  ...
}

The patch will continue auto-incrementing from there.

LLM Accessibility

The site provides llms.txt files for LLM consumption:

• /llms.txt — Structured overview with key links
• /llms-full.txt — Complete documentation as plain text (~1.4 MB)

To regenerate llms-full.txt after content changes:

python3 scripts/build-llms-txt.py

Project Structure

src/content/docs/
├── getting-started/     # Introduction, core concepts, FAQ
├── framework/           # Delegation risk theory
├── applying/            # Principles, examples, tools
│   ├── examples/        # Worked examples
│   └── tools/           # Interactive calculators
├── cross-domain-methods/# Finance, nuclear, mechanism design
├── case-studies/        # Real-world examples
│   ├── ai-systems/      # Sydney, code review bot, etc.
│   ├── human-systems/   # Organizations, nuclear authority
│   └── anomaly-chronicles/  # Fiction series
├── deep-dives/          # Advanced topics
│   ├── theory/          # Theoretical foundations
│   └── research/        # Background research
└── reference/           # Bibliography, roadmap, protocols

License

Content is available under CC BY 4.0.

Contributing

Contributions welcome! Please open an issue to discuss significant changes before submitting a PR.