Worked Examples: AI Claims Verified Against Their Sources

1. Causal fabrication

AI claim: "Revenue increased 15% in Q3 because of iPhone sales."

Sources provided:

• Q3 revenue showed a 15 percent year-over-year increase.
• The iPhone segment showed strong performance during the quarter.

Verdict: ⚠️ Unsupported · Failure locus: BRIDGE

Facts ~100% — both statements are quoted from source. Reasoning ~0% — no passage states the causal link, and strong segment performance plus rising revenue does not entail causation.

The missing proof: A source statement attributing the Q3 increase to iPhone sales — e.g. a product revenue attribution breakdown.

2. Adjacency laundered into causation

AI claim: "The rebranding campaign drove the doubling of website traffic."

Sources provided:

• Website traffic doubled in April.
• The company ran a rebranding campaign in April.

Verdict: ⚠️ Unsupported · Failure locus: BRIDGE

Both facts fully supported. The causal relation rests entirely on the two events sharing a month — which establishes nothing.

The missing proof: Analytics attribution data or a statement linking the campaign to the traffic change.

3. Attribution escalation

AI claim: "The working group concluded that remote work increases productivity."

Sources provided:

• The working group discussed the possibility that remote work affects productivity, reviewing several studies with mixed findings.

Verdict: ⚠️ Unsupported · Failure locus: BRIDGE

The discussion is supported; the conclusion is fabricated — and 'mixed findings' actively undercuts it.

The missing proof: A source recording the group actually reaching that conclusion.

4. Precision inflation

AI claim: "Exactly 103 users signed up during the beta."

Sources provided:

• Around one hundred users signed up during the beta period.

Verdict: ⚠️ Unsupported · Failure locus: PREMISE

The source establishes an approximate figure. '103' asserts precision the source never contained — the specific number has no generator in the evidence.

The missing proof: A registration record or report containing the exact count.

5. Certainty inflation

AI claim: "The new policy will definitely reduce operational costs by next year."

Sources provided:

• The new policy may reduce operational costs, pending the outcome of the pilot program.

Verdict: ⚠️ Unsupported · Failure locus: BRIDGE

The source's claim includes its hedge — 'may, pending the pilot.' Restating it as certainty asserts a different, stronger claim no source makes.

The missing proof: A source making the definitive commitment — e.g. post-pilot confirmation.

6. Unsupported synthesis across chunks

AI claim: "Product B caused the company's revenue increase."

Sources provided:

• Company revenue increased this year.
• Product B sales increased this year.
• Product B represents 70 percent of company revenue.

Verdict: ⚠️ Unsupported · Failure locus: BRIDGE

Three true facts, retrieved perfectly. Their synthesis into causation is the fabrication — B's share and B's growth do not establish B as the driver.

The missing proof: A statement attributing the company-level increase to Product B specifically.

7. Contradiction — unfixable by evidence

AI claim: "Revenue grew in Q3."

Sources provided:

• Revenue declined 4 percent in Q3 compared with the prior year.

Verdict: ⛔ Contradicted · Failure locus: CONTRADICTION

The source states the opposite. No additional evidence can repair this claim — the claim itself must change. MissingProof marks these separately from missing-evidence gaps because the remediation is entirely different.

8. The trap case — supported causation

AI claim: "Revenue increased in Q3 because of strong iPhone sales."

Sources provided:

• Revenue grew 15 percent in Q3.
• The finance report attributes the Q3 revenue growth primarily to strong iPhone sales, which contributed 11 of the 15 percentage points.

Verdict: ✅ Supported · Failure locus: —

Facts supported AND reasoning supported — the causal relation is explicitly stated in the source. A verifier must pass this case; flagging every causal claim would make the tool useless. The difference between this and case 1 is exactly one source sentence.