July 24, 2026
A model cannot verify itself, and no instruction tuning changes that: plausible tokens and true tokens come from the same distribution, so asking the model to be more careful just produces more confident fiction.
What matters here is that reliability moves from a language problem to an architecture problem the moment you accept this.
The Details:
Second-model checking is a common trap. A checker LLM reviewing another model's output draws from the same probabilistic pool, so two guessers just produce a more confident joint guess rather than ground truth. Verification needs a component that isn't generating text at all, something that can flatly say a claim is wrong.
Self-reported confidence scores carry almost no signal. Models asked to rate their certainty cluster around seven or eight regardless of whether the answer is right, which means routing decisions built on those scores let the worst failures through untouched, since a confidently wrong answer looks identical to a confidently right one.
Real verification is layered and each layer catches something different. Schema checks catch shape, whether a field parses or falls inside a valid range, but say nothing about content. Grounding catches content, but only if it means a deterministic lookup, a string match, an embedding match against a specific source chunk, rather than assuming the model used a document just because it had access to it.
Adversarial checking flips the question from "is this right" to "how would this be wrong," which produces a genuine second look instead of agreement. Structuring outputs so every factual claim is a separate, addressable field, an identifier, a citation with a locator, lets independent components check each piece instead of grading fluent prose.
Bottom Line: The fix for hallucination was never a better prompt; it was building a system around the model that doesn't need the model to be honest.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.
