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: 1 November 2026 to 8 November 2026 (UTC).

Every story we publish is split into individual claims and checked by three AI reviewers: Claude (Anthropic), GPT-5.6 Sol (OpenAI) and Kimi (Moonshot AI). Each reviewer marks every claim supported, partial, unsupported or contradicted and must copy a supporting quote from the source material. Our code then checks that the quote exists. Two reviewers must back a claim with a verified quote. This week 211 stories went through the review and 11,795 claims were checked. On 2082 claims (17.7%) the reviewers did not all give the same answer. Here are five of them.

This week in numbers

• 211 stories went through the review (364 review sessions, counting re-checks after corrections); 200 of them are published.
• 11,795 claims checked, 33,874 votes counted (one per reviewer per claim).
• On 2082 claims (17.7%) the reviewers did not all give the same answer.
• Claude (Anthropic): 11,795 votes; backed 96% of the claims it reviewed with a quote we could verify; 0.7% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 424 times.
• GPT-5.6 Sol (OpenAI): 10,235 votes; backed 87% of the claims it reviewed with a quote we could verify; 10.5% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 956 times.
• Kimi (Moonshot AI): 11,592 votes; backed 92% of the claims it reviewed with a quote we could verify; 6.4% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 499 times.
• Gemini (Google): 213 votes; backed 97% of the claims it reviewed with a quote we could verify; 2.8% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 8 times.
• DeepSeek (DeepSeek): 39 votes; backed 100% of the claims it reviewed with a quote we could verify; 0.0% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 0 times.
• Final verdicts: approved 170, sent back for fixes 30, rejected 9, no quorum 2.
• In 3 stories at least one reviewer stepped aside because the story concerned the company that built it.

The cases

Do three reports share a wire origin?

In a World story on a Russian strike on an aid site in Zaporizhzhia, the draft said the reports from three outlets "appear to share a common wire origin." Claude rated the claim partial. It noted that shared wording is visible "but no common wire is identified across all three outlets," and cited a Guardian line crediting Reuters and Agence France-Presse. GPT-5.6 Sol rated it unsupported because no source states or establishes a shared wire origin. Kimi, the dissenting reviewer, rated it supported and quoted an NPR passage about the death toll. That passage does not name a wire service. The claim was published as written: at least two reviewers backed it with a verified quote. The case shows how a hedged inference about sourcing can produce three different readings of the same material.

Story: Russian strike on aid site in Ukraine's Zaporizhzhia kills dozens, reports say

What a federation president actually said

The first draft of a Sports story said the JFA supports Kozo Tashima's FIFA bid "according to directly quoted statements from Tashima and JFA president Tsuneyasu Miyamoto." GPT-5.6 Sol and Kimi rated the claim partial and flagged it. Their note said Miyamoto's quote "does not explicitly state that the JFA backs Tashima." In the flagged quote, Miyamoto calls the decision "a major challenge to contribute to the further development of world football." Claude rated the claim supported. In the next round, the draft attributed JFA support to Tashima's statement and to BBC reporting, which removed the implication that Miyamoto had explicitly stated backing. Other attributions were tightened in the same round. The March election date was attributed to Al Jazeera and The Guardian, and the Rabat ballot and the 211-member total to The Guardian only. The corrected draft was re-checked before publication.

Story: Japan FA honorary president Kozo Tashima declares FIFA presidential run against Gianni Infantino

Whose account, posted where?

The first draft of a story about NASA astronaut Anil Menon voting from space said both reports "attribute the account to Menon's own social media post." Claude and GPT-5.6 Sol rated the claim partial and flagged it. Their note said The Guardian "explicitly ties only the crew remark to a social media post." The flagged passage reads: "He said he filled out the ballot on his station computer and sent it to election officials in Harris county, Texas." Kimi rated the claim supported. The draft was corrected in the next review round. It now says both reports rely on Menon's account, but The Guardian explicitly ties only the crew remark to social media. The corrected draft was re-checked before publication. Who said something and where they said it are separate facts, and each needs its own support.

Story: NASA astronaut Anil Menon casts ballot from space in 2026 US midterm elections

The risk in the word "only"

A Climate story on leaders' plans for a UN climate summit first said the two outlets "agree only on the summit's location in Türkiye." Claude and Kimi rated this sentence supported. GPT-5.6 Sol was the only reviewer to flag it, and it rated the sentence contradicted. Its note said both outlets "also agree that the event is a UN climate summit." It quoted the line "Takaichi to skip U.N. climate summit in Turkey." The draft was corrected in the next review round. The sentence now says the outlets agree that the event is a UN climate summit and that it takes place in Türkiye. The corrected draft was re-checked before publication. A sentence built on "only" makes a strong claim, and a single overlooked point of agreement is enough to make it inaccurate.

Story: Reports: Armenian PM to attend UN climate summit in Türkiye; Takaichi to skip

A brand name, letter by letter

A Health story on the FDA approval of a generic version of lenacapavir first said the brand-name version is sold in the U.S. as "Yeztugo" and is marketed by Gilead Sciences. Claude rated the claim contradicted. Its note said the source spells the name "Yetzugo" and that the claim was not attributed to STAT. It quoted the passage: "The brand-name version, which is known as Yetzugo in the U.S., is marketed by Gilead Sciences." Kimi rated the claim supported. The recorded votes for that round came from Claude and Kimi. In the next review round, the spelling was changed to "Yetzugo" to match the STAT News report. The brand name and the Gilead Sciences marketing claim were also explicitly attributed to STAT News. The corrected draft was re-checked before publication.

Story: FDA approves generic version of twice-yearly HIV-prevention injection lenacapavir

What we learned

In these cases, the difficult calls were about attribution and inference rather than raw facts. The questions were who said something, where they said it, and whether hedged wording such as "appear to" or "may" still makes a claim that needs support. A verified quote shows that the words exist in the source. It does not by itself show that those words support the claim, as the wire-origin case illustrates. The corrections this week were made in the stories themselves: claims were attributed to the specific outlets that reported them. We are watching two things. The first is hedged inferences about sourcing that are published by quorum. The second is cases where matching a source's exact wording, as with the brand-name spelling, is what the check verifies.

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

Traceability
Sources (5): Vocemundi: Russian strike on aid site in Ukraine's Zaporizhzhia kills dozens, reports say · Vocemundi: Japan FA honorary president Kozo Tashima declares FIFA presidential run against Gianni Infantino · Vocemundi: NASA astronaut Anil Menon casts ballot from space in 2026 US midterm elections · Vocemundi: Reports: Armenian PM to attend UN climate summit in Türkiye; Takaichi to skip · Vocemundi: FDA approves generic version of twice-yearly HIV-prevention injection lenacapavir
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: 2026-11-08 15:00 UTC

Produced by the Vocemundi newsroom with AI assistance.