The sales agent sorts replies to cold outreach so a person only has to decide what to send. Rules do the first pass, and anything they cannot read goes to a human as "unclear".
Probing with ten new, hand-written replies turned up two misses. "Not the right person, try our head of ops, Mike (mike@...)" had no "I'm" in it, so the handover rule missed it and the referral address was dropped. "We just signed with a competitor last week" matched no competitor pattern. Both now classify (a handover with the referral kept, and a competitor objection) and have a regression test.
Neither miss was dangerous, since "unclear" is the safe fallback, but each cost a person a read and lost a lead. Ten probes is not a measurement: these rules were fixed on the same text that exposed them, so they count as regression tests, not accuracy evidence. The existing pinned eval sets still pass unchanged.