Review Cluster

Review Cluster — Mid-Draft Adversarial Feedback

Parallel fan-out of a 4-agent read-only core on an active-drafting paper, optionally adding clarity-reviewer as a fifth agent, with auto-synthesise downstream. Lighter than pre-submission-report --parallel; designed for tight iteration, not final-gate verification. Outputs reviews/<scope>/review-cluster/YYYY-MM-DD-cluster-report.md (scope = paper slug from the paper path).

Output Path

Per rules/review-artefact-routing.md (auto-loads in research projects (path-scoped to paper-*/ and paper/)):

• Source slug: review-cluster
• Write reports to: reviews/<scope>/review-cluster/YYYY-MM-DD.md (scope = paper slug) inside the project. Path is relative to the research project root, not the Task-Management repo.
• Never at project root (./CRITIC-REPORT.md-style filenames are forbidden — pre-rule layout).
• Idempotency: if today's file exists, append a same-day descriptor ({date}-revision.md, {date}-r2.md, {date}-pre-submission.md) — never overwrite.
• Index policy: the consolidated cluster report is an output-only artefact and receives no Check=review-cluster row. Each active reviewer receives its own row with Trigger=review-cluster.
• Infrastructure repos (Task-Management, atlas-workspace, etc.): this section does not apply — the path-scoped rule won't load there.

Hard Rules

Existential — block output

1. All active sub-agents are read-only. No git, no latexmk, no edits. See subagent-write-guard.md.
2. Auto-synthesise via synthesise-reviews (unless --no-synthesise). Mid-draft work needs an actionable revision plan, not 4 raw reports.
3. Skip if paper isn't compile-ready — run latex first; review on broken builds is misleading. The skill checks compile-status before dispatching.
4. Cluster is for the user's own papers. For external papers, use peer-reviewer agent instead.
5. Prepare one common review input. After compile preflight, invoke review-packet in fresh-review mode unless the user explicitly supplies a verified packet or requests --no-freeze-input. Record the packet archive hash and canonical PDF hash in every reviewer prompt and in the consolidated report. Packet creation itself creates no reviews/INDEX.md verdict row.

Format — catch in review

6. Write one consolidated cluster report at reviews/<scope>/review-cluster/YYYY-MM-DD-cluster-report.md (scope = paper slug), in addition to the required per-agent reports and INDEX rows.
7. Findings tiered M/m/n (Major / moderate / minor) per severity-gradient.md.
8. Show which sub-agent flagged each finding (audit trail for traceability).
9. Clarity on request. With --clarity (or when the paper's prior referee reviews contain readability complaints), add clarity-reviewer as a 5th parallel agent — reader-experience stall map + the 10 clarity diagnostic classes. Its findings join the synthesis like the others.
10. Yardstick continuity on repeat runs. When a prior cluster report exists for the same paper (reviews/<scope>/review-cluster/), the new run reuses the prior run's frozen yardstick — same phase banner (severity-gradient), same rubric set, and the prior findings list passed to the agents as "verify addressedness + new issues only", never a fresh re-derivation of criteria from the revised text (scores across rounds must be comparable; a moved yardstick fakes improvement). If the prior report is unavailable or the user explicitly resets the phase, mark the report [YARDSTICK-REGENERATED: <reason>] under the phase banner. (Ported from ARS v3.19 re-review protocol, 2026-07-24.)

When to Use

• Active drafting: paper compiles, content is taking shape, want adversarial feedback before final polish
• Mid-revision: addressed first round of supervisor comments; want fresh perspective before next pass
• Pre-pre-submission: ~2 weeks before submission, want to surface major issues with time to address
• Before sharing draft with co-authors

When NOT to Use

• Paper is in final pre-submission state — use pre-submission-report --parallel (13 sub-agents, full kitchen sink)
• Paper is in early scaffold (introduction only, no method/results) — review will flag everything as missing
• Reviewing someone else's paper — use peer-reviewer agent
• R&R revision response — use strategic-revision --external instead (genuine venue-comment-driven)

Modes

Architecture

Phase 1 (preflight) → latex compile check + decision-ready manuscript review packet; abort if broken
Phase 2 (dispatch)  → 4 core read-only sub-agents, plus optional clarity-reviewer, in parallel
Phase 3 (math)      → IF theory paper: verify-math on the model section(s) [orchestrator-run skill]
Phase 4 (consolidate) → synthesise-reviews merges (incl. math verdict) → revision plan
Phase 5 (report)    → reviews/<scope>/review-cluster/YYYY-MM-DD-cluster-report.md

The 4-agent core and optional clarity reviewer

Why these four form the default core:

• referee2-reviewer is excluded — it's the final-stage hostile review, used in pre-submission-report. Mid-draft, hostile review pre-empts genuine improvement.
• artifact-coherence-auditor / reproducibility-auditor are excluded — relevant only when the paper is paired with a replication artifact, which is typically a pre-submission concern.
• code-paper-auditor is excluded — a separate code-suite skill exists for code-side review.
• proofread is excluded — editorial issues are too noisy mid-draft; deferred to pre-submission.

Dispatch contract (evidence-grounded findings)

When dispatching the active reviewer set (Phase 2), each prompt MUST carry the evidence clause from _shared/audit-integrity.md: every finding cites path:line (or §section) AND quotes the exact text it is about, verbatim — no quotable anchor, no finding. Phase 4 (synthesise-reviews) spot-verifies a sample and drops anything it cannot ground, so an agent that emits unanchored findings simply loses them. Tell the agents this up front so they anchor everything.

Phase 1: Pre-flight

# Auto-detect paper or use arg
PAPER_PATH="${1:-$(ls -d paper-*/paper 2>/dev/null | head -1)}"
[ -z "$PAPER_PATH" ] && echo "No paper-*/ directory found" && exit 1

# Check compile-readiness — exit if last latex run failed
LATEST_PDF=$(find "$PAPER_PATH/out" -name "*.pdf" -newer "$PAPER_PATH/main.tex" 2>/dev/null | head -1)
if [ -z "$LATEST_PDF" ]; then
    echo "Paper not compiled or stale. Run latex first."
    # the available structured-question mechanism: run latex now, or proceed anyway (risky)?
fi

Phase 2: Dispatch

Construct the active reviewer set as paper-critic, domain-reviewer, claim-verify, and blindspot; append clarity-reviewer when --clarity is set or prior referee reviews contain readability complaints. Launch every active reviewer in a single parallel dispatch. Each gets:

• Read-only with respect to project files under review — Read, Glob, Grep, Bash (read-only commands only) against the paper / code being reviewed; the agent does NOT modify any project source files

• The standard forbid-list from subagent-write-guard.md

• Paper path explicitly named

• Frozen review identity — the common review-packet path, archive SHA-256, and canonical PDF SHA-256. Reviewers may read the project for anchors, but findings must be compatible with that frozen artifact; if live source has drifted, stop and regenerate the packet.

• Output target — two-step, both required:

  1. Write the per-agent report to reviews/<scope>/<source-slug>/<YYYY-MM-DD-HHMM>.md (scope = paper slug, source-slug = agent name like paper-critic; run mkdir -p reviews/<scope>/<source-slug>/ first), then emit the standard stamp directive. The orchestrator-side propagation step appends the durable INDEX row.
  2. Return a structured findings summary to the orchestrator for the Phase 4 consolidate step.

  These two outputs are NOT mutually exclusive. The file under reviews/<source-slug>/ is the durable artefact and triggers the row stamp; the structured return value is the orchestrator's working copy for consolidation. Earlier wording — "not a file write" — was wrong: it suppressed the per-agent logging step and resulted in 0–1 of 4 cluster dispatches stamping a row in INDEX.md. The 2026-05-17 5-agent patch (commit 23ebcfff) made the agent-side intro unconditional; this dispatch-side fix is the orchestrator-side complement.

Wait for every active reviewer. Do not start Phase 4 (consolidate) until all return.