Trusted Theorem Proving: A Case Study in SLD-Resolution

Abstract

Prolog’s implementation of SLD-resolution furnishes an efficient theorem-proving technique for the Horn-clause subset of first-order logic, and makes for a powerful addition to any automatic or semi-automatic verification system. However, due to the complexity of SLD-resolution, a naive incorporation of a Prolog engine into such a system would inordinately increase the overall trusted base. In this paper we show how to integrate this procedure in a disciplined, trusted manner, by making the Prolog engine justify its results with very simple natural deduction reasoning. In effect, instead of taking SLD-resolution as a primitive inference rule, we express it as a derived inference rule in terms of much simpler rules such as conditional elimination.

This reduction is an example of a general methodology for building highly reliable software systems called certified computation, whereby a program not only produces a result r for a given input x but also proves that r is correct for x. Such a proof can be viewed as a certificate for the result r, and can significantly enhance the latter’s credibility: if we trust the axioms and inference rules used in the proof, we can trust the result. We present a complete implementation of a certifying Prolog interpreter that relies only on three exceptionally simple inference rules: conditional elimination, universal specialization, and conjunction introduction.