adversarial-debate

Adversarial Debate

Three-stage debate pattern for high-stakes decisions:

1. Advocate (Haiku) — strongest case FOR the proposal
2. Critic (Haiku) — strongest case AGAINST, rebutting the advocate
3. Judge (Sonnet) — synthesize and return a structured verdict

Cost ~$0.003 per debate. Takes ~10-15 seconds.

When to use

Good fits:

• Should we merge this PR given tests, blast radius, author tenure?
• Postgres vs DynamoDB for this workload?
• Hire or pass after the last round?
• Ship this feature now or wait a week?
• Trading / investment: enter or stay out?
• Any decision with a defensible case on both sides

Bad fits:

• Questions with one right answer (use a regular call)
• Pure factual extraction (use a cheap single call)
• Open-ended brainstorming (debate narrows, not expands)
• Low-stakes decisions where $0.003 + 15s isn't worth it

Usage

from anthropic import Anthropic
from claude_debate import run_debate, DebateConfig

verdict = run_debate(
    proposal="Merge PR #42: adds Redis cache to auth middleware",
    client=Anthropic(),
    context={
        "diff_size": 340,
        "coverage": "85%",
        "author_tenure": "3 weeks",
        "prod_qps": 2400,
    },
    config=DebateConfig(
        decision_options=("APPROVE", "REQUEST_CHANGES", "DEFER"),
        criteria=["correctness", "rollback cost", "perf under load"],
    ),
)

print(verdict.decision)                    # "APPROVE"
print(verdict.confidence)                  # 72.0
print(verdict.reasoning_summary)           # one sentence
print(verdict.strongest_advocate_point)    # what was most convincing for
print(verdict.strongest_critic_point)      # what was most convincing against

Design choices worth knowing

• Critic sees the advocate's case — so the critic's points are specific rebuttals, not independent arguments. This is what makes the debate adversarial instead of two parallel monologues.
• Judge is explicitly asked for its tipping condition — "what evidence would flip you?" — so the verdict is falsifiable, not just assertive.
• decision_options are bounded — the judge can't invent a new option. Forces a commit.
• Confidence is calibrated in the prompt — 80+ = would bet on it, 50 = genuinely uncertain. Not a vibe.
• Cost asymmetry matches reasoning asymmetry — advocate and critic are structured argument generation (Haiku is fine); judge is reasoning over conflicting evidence (Sonnet).