AI Delegation Curve

A single number tracking what percentage of consequential decisions are made by AI — measured consistently across 9 decision domains.

Live at curve.thinkwright.ai

What This Is

The Keeling Curve measured atmospheric CO2 before anyone cared. This does the same for AI decision-making influence. One composite score (0-100), updated regularly, so the shape of the curve is the argument.

9 domains are tracked: content moderation, algorithmic trading, code generation, customer support, credit decisioning, medical diagnostics, legal AI, hiring, and education.

Each domain score is built from 3-4 normalized indicators sourced from transparency reports, regulatory filings, surveys, and public benchmarks.

Architecture

seed/seed.json          Raw data (manual overrides + collected values)
        |
  cmd/collect           Automated data collection pipeline
  cmd/generate          Transform → Parquet export
        |
  frontend/static/data  8 Parquet files (~27 KB total)
        |
  frontend/             SvelteKit 5 SPA with DuckDB-WASM for in-browser queries
        |
  cmd/server            Go server with embed.FS (single binary, no runtime deps)

Quick Start

# Install dependencies
cd frontend && npm ci && cd ..

# Run the full pipeline: collect → generate Parquet → build frontend
make pipeline

# Or just build and run the server locally
make server
./curve-server -port 8080

Development

# Start frontend dev server (hot reload)
make dev

# Run tests
make test

# Collect fresh data for a single domain
make collect-contentmod

# Full collection + rebuild
make pipeline

Deployment

Deployed to Fly.io as a single distroless container.

make deploy

The Dockerfile runs a 3-stage build: Node (frontend) → Go (embed static assets) → distroless runtime.

Data Pipeline

Collection (cmd/collect): Each domain has a collector that fetches from public sources. Manual overrides in seed/overrides.yaml fill gaps where APIs aren't available, with citations for every value.

Scoring (internal/collect/score.go): Raw values are min-max normalized against historical baselines and theoretical ceilings. Domain scores are weighted averages of their indicators. The composite is a weighted average across domains.

Export (cmd/generate): Transforms seed data into 8 Parquet tables consumed by the frontend via DuckDB-WASM.

License

MIT