메뉴
Skill for enforcing review blinding — ensuring reviewers cannot see each other's reports, managing worktree isolation, handling optional author anonymization, and controlling unblinding at synthesis.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | blinding-protocol |
| description | Skill for enforcing review blinding — ensuring reviewers cannot see each other's reports, managing worktree isolation, handling optional author anonymization, and controlling unblinding at synthesis. |
Use this skill when setting up and enforcing blinding throughout the review process. Blinding is critical for honest, independent review.
The system supports three blinding modes, configured in .review-config.yaml:
| Mode | What's Hidden | Use Case |
|---|---|---|
| Single blind (default) | Reviewers hidden from authors | Standard peer review |
| Double blind | Authors hidden from reviewers + reviewers hidden from authors | Top-tier venues (NeurIPS, ICML) |
| Open | Nothing hidden | Post-publication review |
Blinding is enforced structurally through git worktrees, not through prompt instructions.
# Create isolated worktrees for each reviewer
git worktree add ../review-alpha agent-alpha-<slug>
git worktree add ../review-beta agent-beta-<slug>
git worktree add ../review-gamma agent-gamma-<slug>
Each reviewer's worktree contains:
.review/profiles/reviewer-<role>.yaml)Before dispatching any reviewer, verify isolation:
# Verify reviewer-alpha cannot see beta/gamma reviews
ls ../review-alpha/.review/rounds/round-*/review-beta.yaml 2>/dev/null && echo "BLINDING VIOLATION" || echo "OK"
ls ../review-alpha/.review/rounds/round-*/review-gamma.yaml 2>/dev/null && echo "BLINDING VIOLATION" || echo "OK"
When blinding: "double" is configured:
Strip author identity before distributing to reviewers:
anonymize_submission() {
SUBMISSION="$1"
ANON_DIR="$SUBMISSION/.review/anonymized"
mkdir -p "$ANON_DIR"
# Copy paper with author info removed
if [ -f "$SUBMISSION/paper.tex" ]; then
sed -e '/\\author/,/\\end{author}/d' \
-e '/\\affiliation/d' \
-e '/\\email/d' \
"$SUBMISSION/paper.tex" > "$ANON_DIR/paper.tex"
fi
if [ -f "$SUBMISSION/paper.md" ]; then
# Remove YAML frontmatter author block
sed '/^authors:/,/^[a-z]/{ /^authors:/d; /^ -/d; /^ /d; }' \
"$SUBMISSION/paper.md" > "$ANON_DIR/paper.md"
fi
# Remove cover letter (contains author identity)
# Provide anonymized version or skip
echo "Cover letter withheld for double-blind review" > "$ANON_DIR/cover-letter-note.md"
# Keep metadata but strip author info
yq 'del(.authors)' "$SUBMISSION/metadata.yaml" > "$ANON_DIR/metadata.yaml"
}
| File | Anonymization |
|---|---|
paper.tex/md | Author names, affiliations, emails removed |
metadata.yaml | authors field removed |
cover-letter.md | Withheld entirely |
abstract.txt | Kept as-is (usually anonymous) |
code/ | Kept (may reveal identity — accepted risk) |
When the AE issues a decision, unblinding occurs:
unblind_reviews() {
ROUND="$1"
REVIEW_DIR=".review/rounds/round-$ROUND"
# Make all reviews available
# (In practice, this means the Associate Editor and Revision Agent
# can now read all review files)
echo "Reviews unblinded for round $ROUND"
# Update state
echo "unblinded: true" >> ".review/rounds/round-$ROUND/meta-review.yaml"
}
| Phase | α sees | β sees | γ sees | AE sees | Revision Agent sees |
|---|---|---|---|---|---|
| Phase 1 (Review) | Own work only | Own work only | Own work only | — | — |
| Phase 2 (Synthesis) | Own work only | Own work only | Own work only | All 3 reviews | — |
| Phase 3 (Revision) | — | — | — | — | All reviews + meta |
| Phase 4 (Re-Review) | All reviews + meta | All reviews + meta | All reviews + meta | All re-reviews | — |
Regardless of blinding mode, reviewer identities are ALWAYS hidden:
In .review-config.yaml:
review:
blinding: "single" # single | double | open
Default is single — reviewers are anonymous to authors, but see author names.
Skill providing the exact structured YAML review format that all reviewers must use. Enables automated parsing, scoring, action-item extraction, and convergence detection.
Programmatic validation of figure data and generation code. Since LLMs cannot interpret rendered images, this skill ensures figure correctness by analysing the underlying data and plotting code — catching identical-value plots, missing labels, broken references, and other common defects.
Skill for determining whether the review process has converged (accept), needs another round (iterate), or has reached diminishing returns (exhausted).
Skill for managing the full peer review lifecycle: intake validation, reviewer dispatch, review collection, editorial synthesis, revision cycles, re-review, and convergence determination.
Skill for venue-calibrated scoring, score weighting formulas, decision rules, anti-bias mechanisms, and score interpretation across different venue tiers.
Skill for managing review state persistence — the .review/ directory structure, state.yaml lifecycle tracking, artifact hashing, round management, and final output generation.