synthesise-reviews

Synthesise Reviews

Combine multiple review reports into a single prioritised revision plan with cross-reviewer consensus ranking.

Output Path

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

• Source slug: synthesise-reviews
• Write reports to: reviews/<scope>/synthesise-reviews/YYYY-MM-DD-HHMM.md inside the project, where <scope> is the paper slug (e.g. paper-jtp) or _project for project-level synthesis. 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 update: if reviews/INDEX.md exists, write a one-line entry under "Latest per source" pointing at the new file. Otherwise review-recap will rebuild the index next time it runs.
• Infrastructure repos (Task-Management, atlas-workspace, etc.): this section does not apply — the path-scoped rule won't load there.

Purpose

After running parallel review agents (paper-critic, domain-reviewer, referee2-reviewer), this skill reads all their reports, cross-references issues, and produces a unified REVISION-PLAN.md that groups issues into workstreams by priority and theme.

Inspired by APE Papers' reviewer_response_plan_1.md pattern — workstreams grouped by priority, each concern traced to its reviewer.

When to Use

• After running 2+ review agents on a paper
• After a council review round
• When preparing a revision plan from multiple feedback sources
• After receiving referee reports (processed via strategic-revision)

When NOT to Use

• Before reviews exist — run the review agents first
• To run reviews — use the individual agents (paper-critic, domain-reviewer, referee2-reviewer)
• For a single review — just read the report directly

Workflow

Step 1: Discover Reports

Glob for review files in the project root. Scan the new canonical structure first, then fall back to legacy:

reviews/<scope>/<check>/YYYY-MM-DD*.md (canonical: e.g. reviews/paper-jtp/paper-critic/2026-06-28-1437.md)
reviews/<check>/YYYY-MM-DD*.md (legacy: e.g. reviews/paper-critic/2026-06-28-1437.md)

Where <scope> is a paper slug (e.g. paper-jtp) or _project.

Also check for processed referee comments:

docs/venues/*/revision-*/reviewer-comments/comment-tracker.md

If no reports found, ask the user where the reports are.

Present the discovered reports and their dates. If reports are from different dates, ask whether to synthesise all or just the most recent round.

Step 2: Parse Issues

For each report, extract the issue list:

From paper-critic CRITIC-REPORT.md:

• Parse the Deductions table (columns: #, Issue, Tier, Deduction, Category, Location)
• Extract Critical (C*), Major (M*), Minor (m*) issue details from the detailed sections

From domain-reviewer DOMAIN-REVIEW.md:

• Parse each Lens section's issue table (columns: #, Issue, Severity, Location)
• Map: CRITICAL → Critical, MAJOR → Major, MINOR → Minor

From referee2-reviewer REFEREE2-REPORT.md:

• Parse the structured findings from each audit dimension
• Extract severity-tagged issues

From processed referee comments (comment-tracker.md):

• Parse the comment tracker table
• Each referee comment becomes an issue with severity from the tracker

Step 2.5: Spot-verify findings against their cited location (integrity gate)

Per _shared/audit-integrity.md Rule 2, a review sub-agent's finding is not trusted until its evidence is confirmed — reviewers can emit plausible findings with a fabricated path:line. Before synthesising, spot-verify a random sample of the parsed issues (≥3, or 20% — whichever is larger, weighted toward Critical/Major):

1. For each sampled issue, open its cited Location (path:line) and confirm the quoted text/code is actually there and the issue follows from it.
2. Any sample miss (cited line doesn't exist, quote isn't there, or the claim doesn't follow) ⇒ that report is suspect: widen the check to all of that reviewer's findings and drop every one that can't be grounded.
3. Findings with no Location/quotable anchor at all are dropped, not synthesised — a finding you cannot point at is inadmissible.

Record a one-line Integrity: N sampled, M dropped note in the synthesis output. If reports lack locations entirely and nothing can be verified, say so rather than silently trusting them.

Step 3: Cross-Reference and Consensus Escalation

Match issues across reports by semantic similarity (same underlying problem, possibly described differently):

Important: Consensus can only escalate severity, never reduce it. If one reviewer says Critical and two say Minor, it stays Critical.

Step 4: Group into Workstreams

Cluster issues by theme:

Within each workstream, sort by priority (Critical → Major → Minor).

Step 5: Output Synthesis Report

Write to reviews/<scope>/synthesise-reviews/YYYY-MM-DD-HHMM.md in the project, where <scope> is the paper slug (e.g. paper-jtp) or _project for project-level synthesis. This report contains the consolidated revision plan synthesised from all input reports.

# Revision Plan

**Date:** YYYY-MM-DD
**Reports synthesised:** [list of report files with dates]
**Total issues:** N (C: X, M: Y, m: Z)

## Positive Consensus

Issues/strengths noted positively by multiple reviewers:

- [Strength 1] — noted by [reviewers]
- [Strength 2] — noted by [reviewers]

## Workstream 1: [Theme Name]

| # | Priority | Issue | Flagged by | Action | Source |
|---|----------|-------|------------|--------|--------|
| 1 | Critical | [description] | paper-critic (C1), domain-reviewer (A2), referee2 | [suggested action] | [file:line] |
| 2 | Major | [description] | paper-critic (M3), domain-reviewer (D1) | [suggested action] | [file:line] |
| 3 | Minor | [description] | paper-critic (m2) | [suggested action] | [file:line] |

## Workstream 2: [Theme Name]

[Same table format]

...

## Summary

| Workstream | Critical | Major | Minor | Total |
|------------|----------|-------|-------|-------|
| Identification & Methodology | X | Y | Z | N |
| Mathematical Rigour | X | Y | Z | N |
| ... | | | | |
| **Total** | **X** | **Y** | **Z** | **N** |

## Recommended Order

1. [First workstream to tackle and why]
2. [Second workstream]
3. ...

## Consensus Statistics

- Issues confirmed by all reviewers: N
- Issues confirmed by majority (2/3): N
- Issues from single reviewer: N
- Total unique issues: N

Step 6: Optionally Output REPLY-TO-REVIEWERS.md

If the reviews include processed referee comments (from strategic-revision), also generate:

# Reply to Reviewers

**Paper:** [title]
**Date:** YYYY-MM-DD

## Reviewer 1

### Comment R1-C1: [short title]

> [Original comment — blockquote]

**Response:**
[Placeholder — to be filled by author]

**Changes made:**
- [Placeholder]

### Comment R1-C2: ...

## Reviewer 2

[Same format]

Ask the user before generating this file.

Anti-Patterns

• Do NOT run reviews — only synthesise existing reports
• Do NOT modify source reports — they are read-only inputs
• Do NOT escalate severity beyond consensus rules — if only one reviewer flagged something, keep their severity unless it's confirmed by others
• Do NOT invent issues — only report what the reviewers found
• Do NOT merge issues that are genuinely different — only merge when the same underlying problem is described differently