SpamAssassin as fallback for low-confidence model scores

For messages where the model is uncertain (0.20–0.65 score range), run SpamAssassin and use its score as a tiebreaker. SA has 25 years of tuned rules that catch patterns the model misses.