Abstract—Module extraction can reduce a large rule-based knowledge base, but agreement on returned answers alone does not show whether the same reasoning structure remains available
after compression. ECPM treats extraction as a constrained reduction problem in which answer F1, explanation-family F1, Jensen–Shannon divergence, and counterfactual agreement are
checked together. Evaluation is carried out on three controlled risk workloads and one LUBM-like workload, using three seeds and two bounded reasoning horizons for 24 paired runs. Under the medium explanation granularity used in the main analysis, ECPM-greedy retains a mean module ratio of 0.498, with answer F1 of 0.971, family F1 of 0.904, and counterfactual agreement
of 0.923. Relative to explanation-union, the retained module is smaller by 14.6 percentage points while all configured preservation constraints remain satisfied. Relative to answer-only minimum-support extraction, stronger explanation-family fidelity, lower distributional drift, and higher counterfactual agreement are retained, at the cost of a small reduction in answer recall. The results support bounded explanation-profile preservation as a practical criterion for reusable rule-based modules; no claim is made about global minimality or preservation of every possible explanation.

Knowledge Base Compression, Module Extraction, Explainable AI, Explanation Fidelity, Rule-Based Reasoning, Knowledge Graphs