Certificate-Aware Property-Directed Reachability
For hardware verification practitioners requiring certified results, CAPDR reduces end-to-end cost by shrinking certificates and checker time while maintaining trust.
CAPDR modifies PDR/IC3 to jointly optimize runtime, certificate size, and checker time, using a learned ranking policy over choice points while preserving trust via independent checking. On 2024 HWMCC benchmarks, it solves six more instances than baseline, reduces median certificate-size proxy by 24.6%, and median checker time by 49%.
Property-Directed Reachability (PDR/IC3) is a standard workhorse for hardware safety verification, but most implementations are tuned primarily for time-to-answer and treat the produced invariant or counterexample as a secondary byproduct. In certified workflows, including recent hardware model checking competition rules, the certificate becomes a deliverable whose size, independent checking time, and reproducibility directly affect end-to-end cost. We present CAPDR, a certificate-aware variant of PDR that targets a joint objective over runtime, certificate size, and checker time, while keeping learning outside the trusted computing base. CAPDR exposes a small set of PDR choice points (blocker generalization, obligation ordering, clause pushing, and optional extensions) to a learned ranking policy, but preserves trust by design: every state-changing action is guarded by the same SAT checks as standard PDR, and a SAFE/UNSAFE claim is reported only after an independent checker validates the emitted invariant or trace. We formalize certificate-centric metrics and a replay log that records nondeterministic choices for artifact-grade reproducibility. On the 2024 Hardware Model Checking Competition bit-level safety benchmarks, CAPDR solves six more instances than the baseline. Over each configuration's checker-accepted solved set, the median certificate-size proxy decreases by 24.6% and the median checker time by 49%. Post-fixpoint invariant minimization yields further reductions.