This workflow (RuleSense) supports logic-centered auditing of a rule-based machine learning model — such as rules extracted from a decision tree or random forest, or produced by any other rule-induction method — independently of, or alongside, labeled data. Rules are first extracted from the model as IF-THEN statements and computationally cleaned by detecting and removing contradictory rule pairs (conflicting predictions despite one rule's conditions being more general than the other's) and subsumed rules (redundant rules fully covered by a more general rule with the same outcome). An overview visualization then summarizes, for an analyst-chosen number of discretized value intervals, how each feature and interval is used across rules predicting each outcome, using linked class-wise and feature-wise heatmap matrices with feature-involvement bars. Analysts iteratively apply interactive filters — by outcome, feature inclusion/exclusion, or value interval — to the overview, drilling down into rule subsets to check whether outcome-specific reasoning relies on domain-relevant features and plausible value ranges, flagging any rule groups that omit expected features, use implausible ranges, or otherwise contradict domain logic, and accumulating these into a collection of flagged issues. When labeled data are available, each flagged issue is selected in turn and tested by measuring the impact of removing the implicated rules on classification accuracy, with a human expert deciding — explicitly and reversibly — whether to remove them. To move beyond pairwise feature co-occurrence, rule conditions are discretized into quantile-based intervals, encoded as binary-coded terms, and each rule is treated as a short "document"; Non-negative Matrix Factorization (NMF) is then applied across a range of candidate topic counts, with the number of topics selected by inspecting the distribution of each rule's dominant topic weight across candidates. The resulting topic model is visualized as a topic-term matrix, a per-rule topic-weight line chart colored by predicted outcome, and outcome-specific topic-weight distribution plots, revealing higher-order feature-interaction themes and how well they align with the intended meaning of each outcome. Findings from the overview, filtering, editing, and topic-modeling stages are finally synthesized into explicit statements about the model's logical soundness and domain alignment. The workflow is illustrated generically here using the structure of a case study auditing a random-forest vessel-movement classifier, which exercises the full set of steps (rule cleaning, overview-guided filtering, labeled-data-based impact testing and editing, and topic modeling); the same steps apply, with or without the labeled-data-dependent stage, to rule sets extracted from any other rule-based model.