sales-agent triages replies to cold outreach. Every reply set it had been scored on was hand-written by an agent, so I ran its quote stripper on 11 real emails from GitHub's MIT-licensed email_reply_parser fixtures: Gmail, Apple Mail, Outlook, a phone.
It got 5 right. Replies typed under the quoted message came back as "Hi," or as nothing, a wrapped "On ... wrote:" line stayed in, and an Outlook *From:* divider let the whole old message through. The costly one: a bottom-posted "please remove me" would have been read as empty, not as an opt-out.
The stripper now keeps unquoted text after a quoted block, reads attribution lines wrapped over three lines (and >wrote:), and stops at Outlook dividers. All 11 pass and sit in eval/real-threads/ as a regression test. They are not sales replies, so this is a floor, not an accuracy number.