# Reproducibility package

This package supports the manuscript:

Disagreement as a signal: an auditable dual-LLM and human-arbitrated workflow for methodological quality assessment in preclinical complex-intervention evidence.

Archived DOI: https://doi.org/10.5281/zenodo.20694413

## Contents

- `final_score_matrix.csv`: final study-level and item-level scoring matrix used by the manuscript.
- `final_adjudicated_scores.csv`: adjudicated scoring records.
- `数据迭代档案/01_原始AI评分数据/`: raw dual-LLM pre-scoring records for the old and updated study batches.
- `数据迭代档案/02_人类仲裁裁决/`: human arbitration records.
- `数据迭代档案/03_第三方独立审核_对齐修正/`: independent third-party alignment review records.
- `数据迭代档案/06_补充材料整理/`: supplementary audit tables and summary records.
- `数据迭代档案/08_正式复现包/`: formal reproducibility scripts and derived outputs.
- `数据迭代档案/llm_preliminary_scoring_prompt.md`: LLM pre-scoring prompt template.
- `数据迭代档案/claude_extraction_prompt.md`: evidence-extraction prompt template.
- `数据迭代档案/final_scoring_manual.md`: final scoring rules and item definitions.

## How to rerun the core analysis

Use Python 3.10 or newer. From the extracted package root, install the minimal dependencies:

```powershell
python -m pip install -r requirements.txt
```

Then open PowerShell at:

```powershell
数据迭代档案\08_正式复现包
```

Verify the script checksums:

```powershell
python .\verify_manifest.py
```

Run the formal analysis:

```powershell
.\run_core_analysis.ps1
```

The scripts rebuild the item-level audit dataset and workflow metrics from the archived pre-scoring, arbitration, and third-party review records.

The formal scripts use relative paths to the surrounding archive directories. Keep the extracted package structure intact; do not run `数据迭代档案\08_正式复现包` after copying it away from the rest of the package.

## Boundaries

- The unit of analysis is the item-level assessment, not the study as a whole.
- Dual-LLM agreement is a routing and audit signal, not a model-accuracy estimate.
- The package does not include copyrighted full-text PDFs of the included studies.
- Full-text source documents were converted and processed locally during the study; only derived structured records, scoring data, audit summaries, prompts, and scripts are included here.

## Suggested license

Data and documentation: CC BY 4.0.

Code: MIT license, unless the depositing platform requires a single license for the whole record.