Cross-model verification

A single model's errors are correlated by construction: the same flaw fires on every case that matches it, so mistakes arrive at caseload scale rather than as scattered noise — the mechanism behind the documented automated-benefits failures, where one logic pattern produced tens of thousands of simultaneous wrongful determinations. Human review catches some of this; a genuinely independent second model catches a different slice, because its blind spots are differently placed.

Cross-model verification runs the second model across the first's outputs and treats disagreement as signal: agreements pass, disagreements route to a human. The word doing the work is "independent" — a second instance of the same model, or a second model fine-tuned from the same base, inherits the same blind spots and checks almost nothing. Diversity of vendor, architecture, and training data is what buys de-correlation, and current model-to-model research shows behavior and misalignment can propagate through shared context, so the checking channel itself needs the same provenance discipline as any other pathway.

Two costs are structural, not incidental. First, the deference cost: two systems agreeing reads as proof, and reliance rises on the pair beyond what either earned — the automation-bias literature's finding with a second signature on it. Pair this lever with deskilling arrest and track override rates. Second, the environmental cost: every checked output is inferred twice, making this the compute-heaviest checker in the catalog — the exogenous-cost ledger the framework keeps ("governance is not free") prices it accordingly, and a deployment stacking automated checkers can blow through its oversight-footprint target while the failure rate falls.