Rule-governed behavior and derived relational responding have both been identified as important variables in human learning. Recent developments in the relational frame theory (RFT) have outlined a number of key variables of potential importance when analyzing the dynamics involved in derived relational responding. Recent research has explored the impact of one of these variables, level of derivation, on persistent rule-following and implicated another, coherence, as possibly important. However, no research to date has examined the impact of coherence on persistent rule-following directly. Across two experiments, coherence was manipulated through the systematic use of performance feedback, and its impact was examined on persistent rule-following. A training procedure based on the implicit relational assessment procedure (IRAP) was used to establish novel combinatorially entailed relations that manipulated the feedback provided on the trained relations (A-B and B-C) in Experiment 1, and on the untrained, derived relations (A-C) in Experiment 2. One of these relations was then inserted into the rule for responding on a subsequent contingency-switching match-to-sample (MTS) task to assess rule persistence. While no significant differences were found in Experiment 1, the provision or non-provision of feedback had a significant differential impact on rule-persistence in Experiment 2. Specifically, participants in the Feedback group resurged back to the original rule for significantly more responses after demonstrating contingency-sensitive responding than did the No-Feedback group, after the contingency reversal. The results highlight the subtle complexities that appear to be involved in persistent rule-following in the face of reversed reinforcement contingencies.