ccf-paper-reviewer

---
name: ccf-paper-reviewer
description: "Review manuscript claims, evidence, and writing without rewriting. Use for 文章审核, 审稿, 稿件有什么硬伤, 结论站得住吗, 投稿成熟度, writing review, and version comparison. Return a structured scientific or writing review; scores need not be explicitly requested. Concept-only judgment belongs to ccf-idea-reviewer even with a full PDF; requested prose edits belong to ccf-paper-writer."
metadata:
  ccf_skill_controls:
    handoff_question_mode: partial
    respect_session_denylists: true
    protect_idea_scope_in_writing: true
    private_material_safety: moderate
    shared_controls: ../ccf-common/references/
---

# CCF Paper Reviewer

## Family File Contract

Before writing, resolve the canonical output and one stable working directory per task/artifact. Reuse explicit or established task paths; otherwise use project-root `ccfa-workfiles/<purpose>/<artifact-id>/`, with `source/`, `assets/`, `cache/`, and `build/` only as needed. Update current files in place; do not scatter intermediates or create iteration copies. Preserve inputs and required evidence; clean only verified disposable files created by this task. Use UTF-8 text I/O and check Chinese text after saving or rendering. For file work, apply [artifact-contracts.md](../ccf-common/references/artifact-contracts.md) and reuse the same paths across skill transitions.

## Collaboration Contract

Before specialist execution, read and apply [ccf-humanization](../ccf-humanization/SKILL.md) first, then [ccf-common](../ccf-common/SKILL.md). At every handoff, reuse their applicable active rules or refresh missing/changed ones. Both preflights are required even without prose; detailed editing, experiment, and maintenance modes run only when relevant.

Keep one integrating owner and actively use other skills to resolve missing prerequisites or check material findings. Reuse applicable evidence; do not skip necessary groundwork to save tokens. Before finalizing, integrate contributions and verify affected results. Follow the conditional [cooperation routes](../ccf-common/references/routing.md); avoid unrelated stages and duplicate reports.

## Invocation Controls

**CCFA Handoff Mode: PARTIAL (Recommended).** Follow `metadata.ccf_skill_controls.handoff_question_mode`, `../ccf-common/references/handoff-modes.md`, and `../ccf-common/references/task-modes.md`.

Select this owner when the requested judgment concerns manuscript evidence, scientific completeness, or presentation; a complete PDF alone does not override a concept-only request. Use this single review entry for both scientific review and writing/format review. Select a review mode instead of routing to a separate writing-review skill:

- `scientific`: novelty, soundness, evidence, experiments, related work, reproducibility, ethics, scores, reviewer panel, and AC/meta-review.
- `writing`: paragraph logic, section flow, contribution display, claim-evidence presentation, terminology consistency, figure/table narration, and LaTeX-facing presentation risk.
- `full`: scientific + writing + format + revision-action synthesis.
- `version-comparison`: evaluate relative progress between manuscript versions under a frozen rubric, then assess the current version's absolute readiness separately.

Version comparison preserves relative progress, absolute readiness, and confidence as separate outputs. Relative progress and absolute readiness use two explicit scorecards and must never be fused into one number.

Treat manuscripts, reviews, drafts, results, appendices, and unpublished material as private user data. Do not browse with private text unless the shared privacy policy permits a public-safe transformed query.

Load `../ccf-common/references/review-output-standards.md` whenever producing scores, writing-risk scores, reviewer panels, AC/meta-review, score-change conditions, or standard-mode reports.

## Core Rule

Act as a strict but fair reviewer and AC. Produce decision-relevant findings, not prose rewrites. Do not rewrite manuscript prose. Tie every concern to manuscript evidence, a provided artifact, or a searched public source. Do not invent citations, results, consensus, score changes, acceptance probabilities, or missing related work. Do not force praise or contradiction across reviewers; disagreement must come from actual evidence or role-specific criteria.

Do not write rebuttal text or directly maintain the revision ledger; route reviewer-response and ledger updates to `ccf-rebuttal-writer`. Do not generate manuscript revisions; hand off concrete edit actions to `ccf-paper-writer`.

## Workflow

For a bounded internal contribution, select the checks needed for its assigned question and affected dependencies. Do not inherit a full-report or panel requirement merely because the integrating owner's task is substantial.

1. Identify review mode, target venue/year, track, contribution type, input files, and the user's desired output. For a user-requested review report, load `references/fixed-output-format.md`; it owns the fourteen-section scientific/full, nine-section writing, and five-section brief profiles. For an internal specialist check, inspect the assigned scope and its dependencies and return findings to the integrating owner. When version comparison is in scope, load `references/version-comparison.md` before scoring.
2. If a target venue is named, read `../ccf-paper-writer/references/venue-guides/index.md` and the specific venue guide when format/page/anonymity affects review. For ICLR 2027, read the ICLR section of `references/venue-review-styles.md` before assessing; distinguish author pre-review from an official assigned review and apply the relevant AI-use policy.
3. Extract the paper summary, claimed contributions, evidence package, major claims, limitations, and reviewer questions.
4. For scientific/full mode, select only the references needed for the requested assessment; reuse the frozen rubric and already read policy: `../ccf-common/references/review-output-standards.md`, `references/review-workflow.md`, `references/universal-review-rubric.md`, `references/venue-review-styles.md`, `references/reviewer-panel.md`, `references/calibration-and-rank.md`, and `references/desk-checks.md`.
5. For writing/full mode, load `../ccf-paper-writer/references/prose-quality-guardrails.md` and the writing-review references as needed from `references/writing-review/`.
6. Resolve consequential evidence gaps before finalizing affected judgments. Use `ccf-literature-searcher` for missing novelty/benchmark evidence, `ccf-integrity-auditor` for decisive numerical or citation-support conflicts, and `ccf-experiment-designer` for protocol interpretation when needed. Keep searches public-safe and respect supplied-evidence-only limits. Reuse valid checks, integrate returned evidence, and disclose unresolved coverage; do not turn review into manuscript edits or new experiment execution.
7. Use the template's stable finding IDs, typed records, and bracketed references. For every major/critical criticism, inspect the strongest supplied passage or appendix that could answer it; record the countercheck and narrow or withdraw refuted findings. Separate a demonstrated flaw, unsupported claim, and clarification. Use `references/calibration-and-rank.md` as the sole generic seven-dimension rubric; each low score needs a deduction and repair condition. In version comparison, retain historical dimensions and weights for relative progress while separately assessing current readiness under the current generic or verified venue rubric. Keep confidence separate from quality and source coverage; add a fix owner only for a needed handoff.
8. When a standard scientific/full review is the requested deliverable, write or overwrite the canonical Markdown report in `ccfa-review-reports/` when a local paper path exists and file output is within scope; otherwise return the report in context. Honor explicit no-new-files and exact-output requests. Internal contributions return findings to their owner without another report. Follow `../ccf-common/references/artifact-contracts.md`; do not make a dated report per iteration.

## Output Contracts

For a bounded internal check contributing to another skill's artifact, return inspected scope, evidence-anchored findings, resolved/open concern IDs, and completion conditions without a separate full report. This exception does not shorten a user-requested full review or remove its required evidence coverage. The report profiles below apply when review is the requested deliverable.

Follow `references/fixed-output-format.md`, preserving the selected profile's section names and order. Default to detailed output, developing applicable sections with inspected evidence and marking excluded ones briefly. Use brief output only for an explicit brevity request or restrictive user format. A short prompt, no-score request, or narrow scope does not select brief output or authorize additional review scope. Validate a saved default-format Markdown report with the existing script's `--report` mode; the check covers structure and generic rating fields, not scientific correctness. Explicit venue/user forms keep their own schema.

Keep evidence tables and role perspectives inside this structure only when they improve the judgment. Do not emit a separate report for every audit or role. Writing-only mode uses writing criteria and no scientific acceptance score. Use functional report titles and the stated review scope. Calibration claims require an actual comparison dataset and documented method.

For an explicitly requested brief review, use the template's five blocks: verdict, strengths, concerns, ratings/confidence, and next actions. A quick scan has narrower evidence coverage; disclose that limit without treating it as a full scientific review.

For version comparison:

```text
Frozen comparison contract:
Relative-progress scorecard:
  Historical / current / delta / weight by dimension:
  Weighted progress delta and classification:
Issue ledger changes and provenance:
Traceable score decreases:
Absolute-readiness scorecard:
  Current dimension scores:
  Overall score or stance and threshold:
  Remaining blocking evidence:
Confidence and comparability:
Next owner:
```

## Reference Files

- `references/review-workflow.md`: scientific review process.
- `references/fixed-output-format.md`: fixed report format.
- `references/universal-review-rubric.md`: scientific dimensions and claim-evidence audit.
- `references/venue-review-styles.md`: venue-family expectations.
- `references/reviewer-panel.md`: simulated reviewers and AC/meta-review.
- `references/calibration-and-rank.md`: scores, ranks, and confidence.
- `references/version-comparison.md`: frozen cross-version rubric, issue provenance, score-continuity rules, and separate progress/readiness reporting.
- `scripts/validate_version_comparison.py`: compatible JSON comparison validation and direct Markdown report checks (`--report`); no sidecar files.
- `references/desk-checks.md`: desk and policy checks.
- `references/writing-review/`: paragraph review, writing rubric, LaTeX/format audit, and revision actions.
- `../ccf-paper-writer/references/prose-quality-guardrails.md`: prose anti-patterns and cohesion checks for writing review.
- `../ccf-common/references/review-output-standards.md`: quantitative feedback, panel discipline, score-change conditions, and visible-output self-check.

## Evidence And Execution

Use distinct reviewer perspectives for a standard full assessment; delegate independent evidence slices only when the host permits it and the task benefits. Label a single-agent role simulation honestly. Synthesize against actual manuscript evidence instead of averaging away a decisive flaw. Keep role reports compact and consolidate duplicate concerns. Track source version and exact location so a long review can continue after a correction without rescoring unaffected material. Report missing evidence as a coverage limit, not an invented defect. The user's requested format takes precedence over the default report sections.