--- name: cross-review description: "Run multi-model blind review, challenge, and synthesis." version: 3.0.0 author: Junfu Shi (SJF, xngg1021) license: LicenseRef-Source-Lineage-1.0 platforms: [linux, macos, windows] metadata: hermes: tags: [research, cross-review, multi-model, deliberation, red-team] related_skills: [audit-paper, audit-peer-review, check-submission] --- # Cross-Review Deliberation (`cross-review`) Coordinate **live multi-model blind review, anonymous challenge, and minority-protected synthesis** over a claim, paper, or codebase via `academic-research-kernel` (`3.0.0`) (distinct from `audit-peer-review`, which analyzes historical OpenReview conference records). Replaces legacy `cross-review-five` with a vendor-neutral participant roster (`ReviewPanelSpec`) across CLI and inline MCP. ## When to Use - Running blind multi-model or multi-subagent reviews over a paper, codebase, or technical specification without letting early seats anchor later seats (`--stage plan`). - Clustering structured findings deterministically by `(canonical_target, assertion_key)` across distinct participants, computing complementarity weights, solving $O(N^3)$ Kuhn-Munkres `max_weight_derangement`, and generating anonymized `Proposal-N` challenge bundles (`--stage merge` or MCP `academic_cross_review`). - Dispatching targeted cross-examination only on genuine contradictions and high-severity (`P0`/`P1`) singletons (`--stage challenge`), then reconciling verdicts with **minority-evidence protection** so unrefuted findings backed by verified file locators are preserved in `unresolved-ledger.json` rather than erased by majority vote (`--stage synthesize`). ## Four-Stage Sparse Deliberation Modes - `economy`: `max_challenges=1`, minimal targeted adjudication. - `standard` (default): `max_challenges=3` prioritizing `P0`/`P1` contradictions and singletons. - `audit`: `max_challenges=5` plus maximum-weight derangement peer verification when budget remains. ## Quick CLI Commands ```bash # 1. Define explicit participant roster in panel.json (command templates or offline mock seats) # [{"id": "seat-a", "executor": "command", "provider": "openai", "model": "codex", "cmd": ["codex", "exec", "-o", "{out_path}", "-"]}, # {"id": "seat-b", "executor": "command", "provider": "google", "model": "gemini", "cmd": "gemini -p {prompt_path}"}] # 2. Run step-by-step or end-to-end sparse deliberation academic-research-kernel cross-review ./review-stream --models-file ./panel.json --task ./task.md --stage plan academic-research-kernel cross-review ./review-stream --models-file ./panel.json --stage merge --mode standard academic-research-kernel cross-review ./review-stream --models-file ./panel.json --task ./task.md --stage challenge academic-research-kernel cross-review ./review-stream --stage synthesize academic-research-kernel cross-review ./review-stream --stage status # 3. Compute classic v1 ring-rotation pairings for a given participant list academic-research-kernel cross-review --models "seat-a,seat-b,seat-c" --stage pairings ``` ## MCP Tools - `academic_cross_review` (alias `cross_review_deliberate`): two-step inline deliberation over `findings_by_model`. 1. Call `{"findings_by_model": {...}, "mode": "standard"}` to obtain `registry`, `derangement_pairings`, and anonymized `challenge_plan` (each item includes `challenge_id`, `bundle`, `bundle_sha256`, `run_id`, and `task_sha256`). 2. Call again with SHA-bound reply envelopes `{"findings_by_model": {...}, "challenge_replies": {"C01": {"text": "[CONCEDE] Proposal-1: verified at `target_mod.py:7` where `p < 0.05` fails\n[REFUTED] Proposal-2: counterexample at `target_mod.py:12` preserves `x >= 0`", "bundle_sha256": "<64-hex>"}}}` to synthesize a minority-protected `ReviewRunReceipt`. Strict narrow verdict grammar: optional first line `# Review`, then one single-line verdict per `Proposal-N`; outside single/double backtick code spans (`` `...` ``), characters `[]()<>\\` and `【】` are rejected (wrap expressions like `` `p < 0.05` `` or `` `f(x)` `` in backticks), and the rationale tail must not repeat `[CONCEDE]`/`[REFUTED]`/`[UNRESOLVED]` or the reserved word `proposal` even inside backticks. ## Deliberation Rules 1. Multi-model agreement is a prioritization signal, never proof; provenance locators and deterministic recomputation remain the ground truth. 2. Single-model duplicate findings on the same target stay `singleton` (`consensus` requires $\ge 2$ distinct participants). 3. Challenger selection excludes every participant involved in raising the disputed issue and shuffles proposals as `Proposal-1..N`.