How to Verify an AI Answer Against Source Documents

The manual method, the failure patterns to hunt for, and when to automate.

Whether you're reviewing one high-stakes AI-drafted document or auditing a pipeline's output, verification follows the same logic. Done casually — "read it and see if it looks right" — it fails, because fluent text disarms readers. Done systematically, it is mechanical enough to teach, and to automate.

Step 1: Decompose into claims

Split the answer into atomic claims — single checkable assertions. "Revenue rose 15% in Q3 driven by iPhone sales" is two claims: a fact (revenue rose 15% in Q3) and a relation (iPhone sales drove it). Most verification errors happen right here, by checking the sentence as one unit: the true fact vouches for the fabricated relation riding on it.

Step 2: Check each fact — demand the quote

For every factual claim, find the specific source passage that supports it, and hold the standard at "a reasonable person reading this passage would consider the claim established." If you can't point at a passage, the fact is unsupported, however plausible. Watch for two impostors: paraphrases that shift meaning (survey "responses" becoming "customers"), and numbers with upgraded precision.

Step 3: Check each relation separately

This is the step casual review always skips. List every connective assertion — because, therefore, led to, concluded, will — and for each ask two distinct questions. Is the relation stated anywhere in the sources (a sentence that itself asserts the causal link, the decision, the certainty)? If not, does it necessarily follow from the supported facts? "Both events happened" never entails "one caused the other." "The committee discussed" never entails "the committee decided." Most serious hallucinations in document-grounded AI live in this step and nowhere else.

Step 4: Classify each failure — it determines the fix

• Contradicted: the sources state the opposite. The claim must change; no evidence can be added to fix it.
• Fact unsupported: the data may exist elsewhere — locate the document that contains it, or remove the claim.
• Relation unsupported: the facts stand but the link is invented — weaken the claim to what the facts establish, or find a source that states the link explicitly.

Step 5: Specify the missing evidence

For each fixable gap, write the sentence your sources would need to contain — "the growth was driven primarily by iPhone sales, per the product revenue breakdown" — and name where such a sentence would live and who owns that document. This converts review findings into an actionable work list, and it's the step that makes verification worth the effort.

When to automate

The manual method takes 10–20 minutes per answer done honestly. That's fine for one contract review; it's impossible for a pipeline producing hundreds of answers daily. The method automates faithfully — decomposition, fact checks with quoted spans, separate relation checks, failure classification, and missing-evidence generation — which is exactly what MissingProof runs on every analysis.