---
name: reviewer-rubric
description: 0-100 scoring rubric for peer-reviewer subagent. Defines dimensions, weights, and behavioral indicators for each score range. Loaded only by peer-reviewer.
---

# Reviewer Scoring Rubric

This rubric grounds peer-reviewer's critique in calibrated dimensions. Without it, "review scores" are arbitrary.

## Dimensions

<!-- TODO: 主人按目标 venue 调权重。默认 6 维度，对 ML/computational biology benchmark paper 友好 -->

| Dimension | Weight | What it measures |
|-----------|--------|------------------|
| Novelty | 15% | Is the contribution genuinely new vs. prior work? |
| Technical soundness | 25% | Is the method correct, well-implemented, statistically valid? |
| Experimental rigor | 20% | Are baselines fair, ablations sufficient, evaluation honest? |
| Clarity | 15% | Can a domain expert understand the paper without reverse-engineering? |
| Significance | 15% | Does this matter to the field? Will it be cited / built upon? |
| Reproducibility | 10% | Could a competent researcher reproduce this from the paper alone? |

Overall = weighted sum. Map to recommendation:
- **≥85**: Accept (Spotlight if ≥92)
- **70-84**: Accept (Poster)
- **55-69**: Borderline / Minor Revision
- **40-54**: Major Revision
- **<40**: Reject

## Behavioral Indicators per Score Band

### Novelty (15%)

- **90-100**: Opens a new research direction; reframes a known problem
- **75-89**: Substantial new method or finding within an existing framework
- **60-74**: Meaningful incremental advance
- **40-59**: Combination of known techniques without surprising results
- **<40**: Already published or trivially derivable

### Technical Soundness (25%)

- **90-100**: Method is correct, well-justified, no oversights apparent even on careful reading
- **75-89**: Minor concerns, none load-bearing on conclusions
- **60-74**: Some concerns affect specific claims but not central thesis
- **40-59**: Multiple concerns or one load-bearing concern
- **<40**: Fundamental error invalidating central claims

### Experimental Rigor (20%)

- **90-100**: Strong baselines, sufficient ablations, appropriate statistics, honest about negative results
- **75-89**: Solid evaluation with minor gaps (e.g., one missing baseline, one missing ablation)
- **60-74**: Adequate but with notable gaps reviewers would probe
- **40-59**: Significant gaps: weak baselines, missing ablations, or questionable statistics
- **<40**: Cherry-picked or misleading evaluation

### Clarity (15%)

- **90-100**: Reads cleanly first time, well-structured, appropriate level for venue
- **75-89**: Some sections require effort but no genuine confusion
- **60-74**: Confusing in places; reader has to reverse-engineer intent
- **40-59**: Multiple sections unclear; figures/tables underspecified
- **<40**: Cannot be followed without external help

### Significance (15%)

- **90-100**: Likely to be widely cited; changes how people think about the problem
- **75-89**: Useful contribution; will be cited within subfield
- **60-74**: Useful within a narrow subfield
- **40-59**: Marginal interest beyond the authors
- **<40**: Niche / no clear audience

### Reproducibility (10%)

- **90-100**: Code released, hyperparameters specified, data available, results reproducible
- **75-89**: Most of the above; minor gaps unlikely to block reproduction
- **60-74**: Reasonable effort but some details missing
- **40-59**: Significant gaps; reproduction would require contact with authors
- **<40**: Not reproducible from paper

## Critical Override Conditions

These can drop the overall score regardless of dimension averages:

- **Fabrication suspected**: Score ≤ 30 regardless of other dimensions
- **Clear plagiarism**: Reject, no score
- **Central claim contradicted by own data**: Score ≤ 50 even if other dimensions strong
- **Statistical malpractice (p-hacking, garden of forking paths)**: Cap at 60

## Calibration Notes

- Most submitted papers should fall in 50-75 range. If your reviews consistently score ≥85 or ≤30, you may be miscalibrated.
- The Devil's Advocate Attack should consider whether the strongest hostile reading drops the score by >15 points. If yes, the paper's central thesis is fragile.
- A paper can be technically sound (90 in soundness) but low novelty (40) — these are independent dimensions, do not let them halo each other.