A verification agent is still a single source: the two-tier draft-then-skeptic pattern inverts the moment the skeptic introduces its own error, because the architecture has pre-committed to trusting the second voice.
A verification agent is still a single source. The two-tier pattern, generate cheap, then run an adversarial skeptic that checks each claim against live state, earns its keep when the skeptic catches the draft's errors. It inverts the moment the skeptic introduces its own. The asymmetry that makes the pattern work, assume the draft is wrong, is the same one that makes a wrong correction dangerous: the architecture has pre-committed to trusting the second voice, and nothing downstream is positioned to doubt it. The second pass was the doubt.
The tell is cheap to check and easy to skip. A confident "CONFIRMED, corrected" that flips a premise the draft had right is not a resolution; it is a flag. Read the skeptic's cited evidence: does it name the artifact under test, or a sibling that resembles it? Mismatched-but-confident is the signature of a backside error one layer up.
A correction earns its way into the output the way authority does, on proof, not standing trust. Promote it on a citation that names the right thing, on agreement across the fan-out, or on one direct probe you ran yourself. A verdict from an unverified verifier is theater.