request
wontfix
SpamAssassin as fallback for low-confidence model scores
bill@transcom.net (Transcom IS) ·
2026-05-16 09:50
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.
Comments (1)
bill@transcom.net (Transcom IS)admin
2026-05-16 10:10
Decided against. Adds 100-500ms per scan, conflicts with our "single fast model" architecture. Better to invest in v3 retrain on real production labels.
Sign in to add a comment.