Multi-agent debate method for paper review

[r-uben]

Proposal

Add a progressive_debate review method that uses adversarial multi-agent adjudication instead of monolithic consolidation.

Motivation

The current progressive method has the best raw recall (82.7% location recall on the Refine benchmark, 83% on seeded perturbations). But post-hoc consolidation degrades quality — see #35 for evidence. The consolidation step operates on a JSON blob of comments with no access to the paper text, making blind keep/drop decisions.

Design

Three agents, fixed roles:

Agent	Job
Reviewer	Find issues passage-by-passage (existing progressive pipeline)
Challenger	Attack each candidate: find resolving context, cite conventions, flag duplicates
Editor	See the debate, render verdict: keep / rewrite / merge / drop

Protocol per candidate issue:

Reviewer proposes → Challenger attacks (with paper context) → Reviewer rebuts → Editor decides

Key differences from current consolidation:

• Each debate is per-issue, not one global call on all comments
• Challenger has access to the original passage + running summary
• Debates are independent and parallelizable
• Editor makes local decisions (keep/drop/merge) instead of global pruning

What stays the same:

• Passage splitting, window context, running summary from progressive
• Comment and ReviewResult models (viz/benchmarks still work)
• CLI interface (--method debate)

Scope

• New method_debate.py wrapping progressive's discovery phase
• 4 new prompts: challenge, rebuttal, verdict, merge-cluster
• adjudicate_candidates() replaces consolidate_comments()
• Benchmark against progressive and zero_shot on seeded perturbations

Related

• Consolidation loses 43% of real findings — adversarial adjudication proposal #35 — consolidation quality problem (evidence)
• ROADMAP 260304 #23 — ROADMAP: algorithm design
• PR Add seeded-perturbation benchmark for ground-truth evaluation #34 — seeded-perturbation benchmark