Use Debate vs Ensemble Mode¶

ml-lab offers two review modes. Choose the one that matches your situation.

Debate mode (default)¶

Use when you want a deep, convergent review that resolves every finding to a verdict.

/ml-lab
# When prompted for review_mode, choose: debate

What happens:

1. ml-critic produces initial findings with severity levels
2. ml-defender responds with structured rebuttals (7-type taxonomy: CONCEDE, REBUT-DESIGN, REBUT-SCOPE, REBUT-EVIDENCE, REBUT-IMMATERIAL, DEFER, EXONERATE)
3. Stage B runs 2–4 challenge/response rounds until derive_verdict() determines convergence
4. Each finding gets a deterministic verdict: critique_wins, defense_wins, or empirical_test_agreed

Best for: hypothesis-driven investigations where you want every issue resolved before experimenting.

Ensemble mode (opt-in)¶

Use when you want a high-recall sweep and will triage precision manually.

/ml-lab
# When prompted for review_mode, choose: ensemble

What happens:

1. ml-critic is dispatched 3 times independently — no visibility between critics
2. Findings are clustered by root cause
3. Each issue gets a support count: 3/3, ⅔, or ⅓
4. ENSEMBLE_REVIEW.md is written with tier-weighted output

Best for: broad exploratory sweeps where missing a real issue is costlier than triaging false positives. ⅓ minority findings need explicit user confirmation before entering experiment design.

Choosing between them¶