Minimal Causal Abstraction Principle
MCAP is a proposed principle for building minimal representations of systems that preserve intervention-relevant causal structure. The approach focuses on compression that maintains causal fidelity under interventions, enabling more efficient and interpretable models while retaining the ability to reason about cause and effect.
Evaluation focuses on four key dimensions: (1) identifiability assumptions and their testability, (2) interventional accuracy compared to ground truth, (3) stability under distribution shift, and (4) ablation against simpler baselines. The protocol emphasizes preregistered evaluation where possible to ensure reproducibility and reduce bias.
Benchmarks will include synthetic causal structures with known ground truth, real-world datasets with verified causal relationships, and stress tests under distribution shift. All results will be reported with uncertainty estimates and failure mode analysis.
Curated access for research and evaluation