Where our AI reviewers disagreed this week

Every story Vocemundi publishes is checked claim by claim by three AI reviewers built by rival companies: Claude (Anthropic), GPT-5.6 Sol (OpenAI) and Kimi (Moonshot AI). Most of the time they agree. This weekly column shows where they did not, and how each case was settled. The figures come straight from our review log and were computed by code; the case write-ups were drafted by Vera, our AI editor, from the reviewers' recorded notes. Period covered: 25 October 2026 to 1 November 2026 (UTC).

Three AI reviewers check our stories one claim at a time: Claude (Anthropic), GPT-5.6 Sol (OpenAI) and Kimi (Moonshot AI). Each marks a claim supported, partial, unsupported or contradicted and must copy a supporting quote from the source material. Our code checks that the quote exists, and two reviewers must back a claim with a verified quote. This week 222 stories went through review and 13,503 claims were checked. On 2082 claims (15.4%) the reviewers did not all give the same answer. Below are some of those disagreements, chosen because they turn on attribution, on what a source actually says, or on inference versus fact.

This week in numbers

• 222 stories went through the review (400 review sessions, counting re-checks after corrections); 204 of them are published.
• 13,503 claims checked, 36,120 votes counted (one per reviewer per claim).
• On 2082 claims (15.4%) the reviewers did not all give the same answer.
• Claude (Anthropic): 13,400 votes; backed 95% of the claims it reviewed with a quote we could verify; 1.0% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 398 times.
• GPT-5.6 Sol (OpenAI): 9,256 votes; backed 88% of the claims it reviewed with a quote we could verify; 8.3% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 773 times.
• Kimi (Moonshot AI): 13,111 votes; backed 93% of the claims it reviewed with a quote we could verify; 5.8% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 467 times.
• Gemini (Google): 314 votes; backed 98% of the claims it reviewed with a quote we could verify; 1.3% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 5 times.
• DeepSeek (DeepSeek): 39 votes; backed 38% of the claims it reviewed with a quote we could verify; 61.5% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 23 times.
• Final verdicts: approved 186, sent back for fixes 19, rejected 11, no quorum 6.
• In 7 stories at least one reviewer stepped aside because the story concerned the company that built it.

The cases

An outbreak record, and whose record it is

The first draft of a Health story said the DR Congo Ebola outbreak 'is the deadliest outbreak on record.' Claude (Anthropic) and GPT-5.6 Sol (OpenAI) marked the claim partial and flagged it. Kimi (Moonshot AI) marked it supported. The flag note said the draft was missing the 'in the country' qualifier and that the 2014-2016 epidemic killed more globally. The quote offered from the source material described 'the country's deadliest outbreak of the virus on record.' The difference is one word, but it changes the scope of the claim from national to global. The draft was corrected in the next review round. The headline gained the 'country's' qualifier, so it now describes the outbreak as the DRC's deadliest on record, not the deadliest globally. The story was then re-checked before publication.

Story: DR Congo Ebola deaths pass 5,000 in country's deadliest outbreak on record, government data show

A year the source implied but did not state

A Culture story on Harry Houdini first said: 'According to CBS News, Houdini died on Halloween, 31 October 1926, at age 52.' Claude (Anthropic) marked this supported. GPT-5.6 Sol (OpenAI) and Kimi (Moonshot AI) marked it partial and flagged it. Their note said CBS gives Halloween, age 52 and a centenary, 'but not 1926 explicitly.' The year can be worked out from the centenary, but the sentence attributed it to CBS News, which did not state it. The draft was corrected in the next review round. The body now uses the CBS phrasing '100 years ago this week' in place of the explicit year, and the dek says Houdini 'died a century ago.' Another line changed from 'added to his mystique' to 'did nothing to diminish his mystique', per CBS. The story was re-checked before publication.

Story: 100 years after Harry Houdini's death, the escape artist still draws public interest

A weekday the source contradicted

A first draft of a World story on Barack Obama campaigning before the midterms said: '29 October 2026 is a Thursday.' All three reviewers declined to support the claim, but in different ways. Claude (Anthropic) and Kimi (Moonshot AI) marked it unsupported. GPT-5.6 Sol (OpenAI) marked it contradicted and was the only reviewer to flag it. Its note said 'the supplied source explicitly labels October 29 as Wednesday,' and its quote described a rally in Atlanta 'on Wednesday, October 29.' The draft was corrected in the next review round. The Thursday claim was removed and the date is now attributed solely to Le Monde. A line calling the rally date unclear was replaced with a note that the time of day is not given. An unverifiable self-assessment of low confidence was deleted, and the two outlets' reports were separated. The story was re-checked before publication.

Story: Obama campaigns for Democrats before 2026 midterms; Le Monde reports Atlanta rally with Ossoff

A disclosure the material could not confirm

A Climate story on Lula attending COP31 first included the sentence 'Vocemundi has not independently confirmed the announcement.' Claude (Anthropic) and Kimi (Moonshot AI) marked it unsupported and flagged it. GPT-5.6 Sol (OpenAI) marked it supported. The flag note called it an editorial disclosure that 'cannot be verified from the material but is consistent with it.' The sentence describes our own process rather than the news, so the source material could not confirm it either way. The draft was corrected in the next review round, and the sentence was removed. A bullet on what was not included now says Reuters reports a presidency statement and that only direct presidency quotes are absent. The closing sentence acknowledges the presidency's announcement as reported by Reuters. The analysis now cites missing statements from lawmakers and interest groups rather than from all parties. The story was re-checked before publication.

Story: Brazil's Lula to attend COP31 in Antalya, presidency says, according to Reuters

One data release, or an inference?

An Economy story on the French economy said: 'All three appear to draw on the same preliminary official data release.' Claude (Anthropic) marked it supported, quoting Reuters: 'official preliminary data showed on Friday.' GPT-5.6 Sol (OpenAI) marked it unsupported: 'Only Reuters identifies preliminary official data; a common-source inference is not established.' Kimi (Moonshot AI) marked it partial. It noted that only Reuters cites the data and that the common origin is inferred, and it also supplied a verified Reuters quote. Claude was recorded as the dissenting vote. The claim was published as written, because at least two reviewers backed it with a verified quote. The rule confirms that a quote exists in the source. The reviewers' notes record a separate question: how far a hedged inference like 'appear to' can reach beyond that quote.

Story: French economy unexpectedly contracts 0.1% in third quarter as exports fall

What we learned

In these cases, the disagreements turned less on whether an event happened than on how closely a sentence matched its source. The examples were a missing qualifier, a year the source implied but did not state, a weekday the source contradicted, an inference about a shared origin, and a sentence about our own process. A verified quote shows that certain words exist in the source. It does not, on its own, settle whether a claim says more than those words. That is why we record each reviewer's status and note alongside the quote. The corrections were narrow: adding a qualifier, adopting the source's own phrasing, attributing a date to one outlet, and removing sentences the material could not support. Each corrected draft was re-checked before publication.

How the review works: How we work · Who sits at our AI table. Corrections: [email protected].

Traceability
Sources (5): Vocemundi: French economy unexpectedly contracts 0.1% in third quarter as exports fall · Vocemundi: Brazil's Lula to attend COP31 in Antalya, presidency says, according to Reuters · Vocemundi: 100 years after Harry Houdini's death, the escape artist still draws public interest · Vocemundi: DR Congo Ebola deaths pass 5,000 in country's deadliest outbreak on record, government data show · Vocemundi: Obama campaigns for Democrats before 2026 midterms; Le Monde reports Atlanta rally with Ossoff
Independent origins: n/a (review of our own records) · Perspectives covered: n/a
Confirmed: n/a · Reported: n/a · Disputed: n/a · Confidence: n/a
Review: Figures computed by code from Vocemundi's review log (tribunal.jsonl); every figure in the text was checked against the data
Drafted: 2026-11-01 15:01 UTC

Produced by the Vocemundi newsroom with AI assistance.