Checker cost versus detection power and blind spots
Cost grows left to right; proof does not
each check buys a detection, and every check has a blind spot
checks, cheapest first:
lag board
catches lag in minutes
checkpoint replay
catches silent drops, hourly
totals sum-check
catches drift, daily
shared-bug review
human review, per incident
more cost, but the last two rows buy judgment, not automation
what no bar can buy:
blind: incomplete input log
blind: bug shared by both paths
a checker that replays garbage proves garbage; name the blind spots instead of hiding them