v0.1.0 — Prove the Loop

AdversarialDebate v0.1.0 — Independent dual-review pass with bounded adversarial debate.

Field Test Results

• 411 debates across 70 real PRs (kubernetes, prometheus, golang/go, etcd, rails, django)
• 6 model pairs × 4 models — total cost: $0.53
• Binary bar PASSED: 49/49 PRs with known outcomes had ≥1 debate claim matching the actual cause
• Theater rate: 0.2%, Verdict stability: 96%, Engine errors: 0

Key Findings

• Model diversity is the strongest predictor of productive debate
• The debate prompt is the critical path, not the engine
• Homogeneous (GPT+GPT) outperforms weak diversity (GPT+Gemini)

Install

pip install adversarial-debate

See CHANGELOG.md for full details.