Explanations Using Defeasible Rule Sets

Symbolic classifiers that reason by default, retracting a conclusion only when a more specific exception applies, offer predictions that are interpretable by construction rather than through post-hoc justification. Exception Closed Conjunctive Rule Sets (ECCRS) realize this idea directly: classification is decided by a Most Specific Wins(MSW) operator, so the explanation for any prediction is simply the chain of rules and exceptions that produced it. This structural account is easy to read, but it does not on its own show that the decision is justified under an established theory of reasoning with exceptions. We build a workbench that checks ECCRS’s decisions against such a theory. Building on a correspondence established by Madzime and Meyer, it translates an ECCRS rule set into the equivalent form used in the KLM framework, a well-established formal account of how conclusions should be drawn and withdrawn in the presence
of exceptions, and tests whether three of its defeasible reasoning methods, namely  Rational Closure (RC), Lexicographic Closure (LC) and System W (SW),reaches the same verdicts as ECCRS. The Alignment Theorem states the three assumptions that ECCRS rules must have for 100% alignment, namely  Sanity, Strictly Global Exception Closure and No total Override. On rule sets meeting these conditions, our workbench records a 100%  alignment between ECCRS and the three defeasible reasoning methods on every tested instance. We then run the workbench 
on rule sets that break those conditions, and find that disagreement arises only when a condition is violated, and never as a wrong label: the formal reasoner withholds a decision or reports a contradiction rather than contradicting ECCRS’s answer. Comparing the two sides’ explanations, we show that the formal account can support and supplement ECCRS’s own rule-based justifications