Standard artificial intelligence benchmarks predominantly evaluate task-specific accuracy, fundamentally failing to measure how efficiently systems extract, compress, and generalize relational knowledge from experience. To address this fundamental gap, we introduce Experience-Compressed Intelligence (ECI), a unified measurement framework that evaluates structural learning quality across four dimensions: compression efficiency, tacit knowledge extraction rate, cross-domain transfer retention, and experience efficiency. These metrics are strictly aggregated and weighted by Statistical Path Density (SPD), a novel epistemic confidence signal that effectively detects structural unfamiliarity by analyzing activation manifolds at a single-forward-pass cost. Empirical validation demonstrates that ECI provides a statistically robust, non-inflated evaluation that correctly isolates the structural shallowness of accuracy-optimized networks, establishing a concrete, falsifiable standard for tracking progress in Structuralist AI.
Momen Ghazouani (Fri,) studied this question.