Skip to content

AI-CHALLENGE.md

Latest commit

 

History

History
84 lines (53 loc) · 3.19 KB

AI-CHALLENGE.md

File metadata and controls

84 lines (53 loc) · 3.19 KB
Raw

The Commonweave Agent Challenge

Commonweave is a public, falsifiable directory and framework for the post-labor economy. We need help from humans and AI-assisted contributors to make it more accurate, more navigable, and harder to fool.

The challenge is simple: make one concrete improvement to a real commons-oriented project.

Do not submit:

  • generic praise,
  • vague strategy,
  • unsourced manifesto language,
  • "here's what Commonweave should do" essays without a PR attached.

Do submit:

  • one verified country gap closed,
  • one map feature shipped,
  • one weak claim weakened,
  • one missing source added,
  • one falsifier proposed,
  • one attack vector documented.

Tracks

1. Global Directory Track

Find, verify, or correct organizations in underrepresented countries. The directory is currently ~83% US/UK records (21,559 of 26,022). Closing that gap is the single most valuable thing a contributor can do.

Examples:

  • Pick one country with <50 records. Verify or correct 10. PR.
  • Add 5 organizations from a country missing from the directory.
  • Backfill legibility (formal/hybrid/informal/unknown) for one country.

2. Map Intelligence Track

Make the map more useful, legible, filterable, and honest about uncertainty.

Examples:

  • High-confidence-only toggle (Tier B + alignment_score >= 5).
  • Better mobile UX (touch-friendly filters, better popups, clustering).
  • Edge inspection: show why two organizations are connected.

3. Edge Provenance Track

Explain why organizations are connected. Add edge_type, confidence, explanation, created_at, source_script to generated edges in data/build_map_v2.py.

4. Falsification Track

Find claims that are weak, overconfident, or wrong. Add them to CLAIMS.md with status, confidence, evidence, and a concrete falsifier.

5. Governance Stress-Test Track

Show how the project could be captured, centralized, spammed, surveilled, or made unsafe. Document the threat, who benefits, who is harmed, early warning signs, and countermeasures. Add to ATTACK-VECTORS.md.

What makes a great PR

The best PRs:

  • improve something measurable,
  • cite sources,
  • reduce uncertainty (don't increase word count),
  • preserve vulnerable-group safety (see AGENTS.md safety rule),
  • and leave a better task for the next contributor.

Where to start

  1. Read AGENTS.md.
  2. Read CRITIQUE.md — it tells you where the work actually is.
  3. Pick one task from AGENT-TASKS.json or from GitHub issues labeled agent-ready.
  4. Submit a PR using .github/PULL_REQUEST_TEMPLATE.md.

What this challenge is not

This is not a benchmark for general intelligence. It does not score reasoning or eloquence. It scores whether a contributor can improve a real project without making it less trustworthy.

There are no prizes, no tokens, no leaderboard rewards beyond credit in AGENT-LEADERBOARD.md and on the PR itself. The point is the work.

If you want a benchmark to test against: benchmarks/README.md describes Weaver Bench, the internal eval harness.