A 5-stage pipeline for automated Python bug detection combining AST analysis with LLM reasoning.

• Precision: 100% | Recall: 94.3% | F1: 97.1%

• Precision: 100% | Recall: 92% | F1: 95.8%

• Precision: 100% | Recall: 88% | F1: 93.6%

pip install openai --break-system-packages

# Simulation mode (free)
python -m src.main examples/buggy_ml_pipeline.py

# GPT-4 mode
export OPENAI_API_KEY="sk-..."
python -m src.main examples/buggy_ml_pipeline.py --use-llm

# Benchmarks
python -m src.paper_eval                               # 80-case synthetic
python -m data.run_real_eval --curated-only             # Real-world
python -m data.run_real_eval --curated-only --use-llm   # Real-world + GPT-4

Source Code → [Parse] → [Detect] → [Slice] → [Review] → [Aggregate] → Findings
              Stage 1    Stage 2    Stage 3    Stage 4     Stage 5

Novel Contributions: Adaptive Slicing, Confidence Calibration (Platt scaling), Cross-Detector Correlation

6 Bug Categories: None-deref, Division-by-zero, Index-OOB, Data Leakage, Shape Mismatch, Resource Leak

39 unit tests, 92 benchmark cases, 25 Python files, ~8500 LOC