A toolkit for systematically checking the proofs in statistics / ML-theory paper appendices using coding agents (Claude Code, Codex, etc.).

Authors: Wei Ma

proofcheck-stat-paper/
  README.md       ← you are here
  guideline.md    ← principles, templates, severity system
  workflow.md     ← Pass 0-5 execution methodology
  bootstrap.md    ← prompt to auto-generate CHECK_PLAN + EXECUTION_ORDER
  scripts/        ← automation (index, crossref, scaffold)
  templates/      ← PROGRESS.md and other reusable templates
  tutorial/       ← step-by-step walkthrough (CBARA paper)
  papers/         ← put papers to check here
    your-paper/   ← one folder per paper
      audit/      ← created during check execution

Everything happens inside this directory. Open Claude Code here.

1. Create a folder: papers/your-paper-name/
2. Put the paper's LaTeX source into it: papers/your-paper-name/paper.tex
3. Tell Claude:

Check the appendix proofs of the paper in papers/your-paper-name/paper.tex.
First read guideline.md and workflow.md to understand the methodology,
then read the paper and run bootstrap.md to generate a check plan.

Claude will:

• Generate CHECK_PLAN.md and EXECUTION_ORDER.md inside papers/your-paper-name/
• Create an audit/ directory there
• Execute the phases (indexing → foundations → lemmas → theorems → final report)

All outputs stay in papers/your-paper-name/. The toolkit files at the root are never modified.

Tell Claude: "The paper in papers/your-paper-name/ has been updated. Re-run the checks."

Read this first. Paper-agnostic. Defines:

• Core objective, operating principles
• Severity system: S0 (fatal) through S3 (minor)
• Proof-unit templates, 19 common failure patterns
• Agent prompt templates for every stage

The bridge from principles to execution. Covers:

• How to extract proof architecture from a paper
• How to build dependency-ordered execution plans
• Workspace setup, phase-by-phase execution
• Paper-type-specific adaptations (asymptotic, concentration, optimization, Markov chain, M-estimation)

One prompt → CHECK_PLAN.md + EXECUTION_ORDER.md drafts.

• Mode A: fully automatic (best when paper has clear proof-strategy section)
• Mode B: you provide a knowledge block with proof strategy insights

A step-by-step walkthrough using the CBARA paper. Shows what every file looks like at every stage — not a paper to check, but a reference for how the toolkit works.

• Claude Code (or similar coding agent with file read/write)
• LaTeX source of the paper
• Git (for version-tracking toolkit changes)

Tested with Claude Code using deepseek-v4-pro at maximum effort. Results may vary with other models or effort settings.