When organisations "fact-check" AI output, they typically send the same text back to the same model (or a model from the same provider family) and ask "Is this correct?" This creates a reliability illusion.
The self-agreement problem
A large language model asked to verify its own output is structurally biased toward agreement. It shares the same training data, the same architectural limitations, and often the same failure modes.
Single-model verification fails because:
- Same training bias — The verifier shares the same data blind spots as the generator.
- Correlation not independence — Errors in generation correlate with errors in verification when both come from the same source.
- Confidence is not accuracy — Models express high confidence even when they are wrong.
- No adversarial tension — Without genuine disagreement, the process is confirmation, not verification.
The VerifiedOS approach
VerifiedOS uses four independent model routes from different provider families. The primary model assesses the claim. Three heterogeneous judge models independently evaluate the same evidence bundle.
Crucially, the adjudicator that determines the final outcome is not another generative model — it is deterministic application logic that applies versioned outcome rules.
Why this matters
If all four routes agree (SUPPORT), and all evidence is grounded, the result is VERIFIED. If any route dissents with valid reasoning, the result is CHALLENGED. If any required component fails, the result is UNVERIFIED.
Disagreement is preserved, not hidden. That is the difference between verification and confirmation.