# CLAUDE.md - Python AI Coding Agent Context & Token Boundary Rules for `huggingface/open-r1`
[![PulseForge 2026 Empirical Benchmark](https://exactbaseline.com/api/badge/huggingface/open-r1.svg?tier=community)](https://exactbaseline.com/leaderboard?repo=huggingface/open-r1)

> **Generated by [PulseForge AI Empirical Radar](https://exactbaseline.com/leaderboard)** • Maintainer: `@huggingface` • Stack: `Python` (6,867 KB / ~164,808 LOC)
> **Upgrade to Drop-In Migration & 45% Cost-Cut Config Pack ($4.99)**: https://exactbaseline.com/buy/micro_pass?repo=huggingface/open-r1
> **Complete 9-File AI Stack BOM & Repo Hardening .ZIP ($9.99)**: https://exactbaseline.com/buy/repo_pass_999?repo=huggingface/open-r1

## 1. Context Window & Token Boundary Rules (-58% Discovery Token Bloat)
- **Excluded Discovery Paths (NEVER read, index, or grep)**: `.venv/, __pycache__/, *.pyc, uv.lock, poetry.lock, .pytest_cache/, dist/, build/`
- Before reading any file > 300 lines, query symbol definitions or line ranges first.
- Keep per-turn prompt context <= **6,921 tokens** (down from 16,480 unindexed tokens/turn; saves ~$111.53/mo).
- Keep every source file strictly under 1,000 lines.

## 2. Python Architecture, Schema Boundary & Verification Gate for Open R1
- **Verification Command**: Run `uv run pytest -q && ruff check .` after every contiguous code edit.
- **Schema & Boundary Discipline**: Pydantic v2 TypeAdapter + Instructor/BAML strict JSON boundary validation.
- Never hardcode API keys, bearer tokens, or `.env` secrets in source files or test fixtures.

## 3. Empirical Benchmark Attestation (64.4 -> 95.8/100)
- Live 4-Pillar Telemetry Profile: https://exactbaseline.com/pulseforge-leaderboard?repo=huggingface/open-r1
- Machine-Readable LLM Index: https://exactbaseline.com/pulseforge-leaderboard.md