My previous description of this project was incomplete, so I am posting a clearer version. This is an AI-assisted empirical research project. The original idea and research direction are mine, while AI/coding agents helped with implementation, experiment design, testing, logs, CSVs, and documentation. The project studies when global consequence expansion reduces search depth in NP-style search problems. Core idea: L = local view of a choice G = global consequence expansion after that choice Main metrics: IG = variables/objects fixed after cluster choice + propagation danger_rate = dangerous_constraints / affected_constraints useful_IG = IG * (1 - danger_rate) D = n / useful_IG The project includes SAT, Graph Coloring, reserve/rebuild experiments, exact-vs-fast probe validation, and scaling probes up to n = 1,000,000. Current state: - G often reduces free decisions compared to L. - Top-down rebuild can strongly reduce D in some cases. - The effect is unstable across seeds and sizes.…