axelrod-review

Axelrod Review — Game Theory for AI Review Loops

AI reviews are stateless. Same prompt, same depth, every time. Your reviewer doesn't know it flagged the same issue last round. It doesn't know you fixed everything it asked for. It treats every draft like it's meeting you for the first time.

This skill gives your AI reviewer memory and consequences — using the same tit-for-tat strategy that won Robert Axelrod's famous computer tournament in 1984.

The Problem

You run an AI review on your draft. It flags 5 issues. You fix 3 of them (the other 2 were bad suggestions). Next review? Same depth. Same scrutiny. No reward for cooperation.

Worse: if you ignore feedback consistently, nothing changes either. No escalation. No consequences. The reviewer is a broken feedback loop.

The Fix: Tit-for-Tat Review Dynamics

Track a cooperation score between you (the author) and the reviewer. The score adjusts review depth over time.

How it works

cooperation_score = issues_addressed / (issues_found − rejected_with_reason)
                    (rolling avg, last 5 rounds)

Every flag gets one of three responses, each priced differently:

• Fix it → cooperation
• Reject it WITH a one-line logged reason → also cooperation (a wrong flag deserves a documented "no", not silent compliance)
• Ignore it silently → defection. Visible in the ledger immediately.

This is the key incentive: rejecting a bad flag costs one written line. Ignoring it costs trust. The cheap path and the honest path become the same path.

• Score > 0.8 (high trust): Light review — hook, accuracy, one standout issue only
• Score 0.5–0.8 (standard): Full review — all checks, all fixes
• Score < 0.5 (low trust): Deep audit — recurring issue tracking, verify every claim, no skimming

Setup

1. Create the trust ledger

Create review-trust.json in your project:

{
  "meta": {
    "strategy": "tit-for-tat",
    "description": "Cooperation tracking for AI review loop",
    "started": "2026-01-01"
  },
  "rounds": []
}