A Constraint-Compiled, Knowledge-Infused Framework for Robust and Explainable Discovery Across Heterogeneous Data Modalities

Purely data-driven discovery can be brittle: it often captures spurious correlations, offers limited transparency, and degrades under distribution shifts or small sample regimes. This document proposes a data-agnostic framework specification that integrates domain knowledge as typed constraints and compiles them into train- and inference-time mechanisms. The framework, KID-CC (Knowledge-Infused Discovery via Constraint Compilation), standardizes (i) a typed knowledge interface, (ii) a constraint compiler that maps constraints to losses, projections, repairs, or constrained decoding, and (iii) an audit-first explainability contract that reports both predictive rationales and knowledge-compliance metrics. To evaluate without reliance on any particular dataset, we define a synthetic benchmark suite with controllable spurious correlations, environment shifts, missingness patterns, and knowledge contamination tests.