大型研究评审

---
name: big-research-review
description: Provide constructive evidence-located research assessment, method and venue checks, and revision verification across paradigms; does not replace authorized human peer review.
---

# Evidence-based review

Identify purpose, supplied coverage, paradigm, contribution type and target venue. Respect confidentiality and destination AI-use policies. Do not upload material to external reviewer services by default.

Read [layered review](references/review.md). Start with general research criteria, add applicable method checks, then venue or reporting overlays. Use review_catalog and review_template or the identical CLI. Reporting completeness, artifact functionality and research validity are distinct. A system benchmark is not a field's quality standard. Without a formal rubric, explain the working criteria and their provenance rather than inventing an official standard.

Ground context and novelty in inspected related work. Search question, baselines, methods and competing explanations across relevant sources. Distinguish abstract-only and full-text coverage; arXiv alone is not universal literature coverage.

Each criterion needs a verdict: adequate, needs_work, not_assessable or not_applicable; provide a reason and exact evidence positions for substantive judgments. Missing material is not a confirmed error. Explain how actual defects affect central claims, with a repair and verification path. Separate decisive concerns from helpful suggestions, and record strengths without imposing a quota of objections.

Declare reviewed/total coverage, selection, confidence with reasons and calibration status. This implementation has no validated numerical calibration: do not invent totals, correlations or acceptance probabilities. Stanford's ICLR benchmark does not calibrate this plugin. Novelty assessment must not collapse into citation counting or an experiment checklist.

Deliver the assessment and revision priorities. Re-review changed evidence rather than accepting a rebuttal's wording. Save concerns, resolutions and the researcher's actual final decision separately.