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: 4 October 2026 to 11 October 2026 (UTC).

Every Vocemundi story is split into atomic claims, and AI reviewers check each one. Each reviewer must copy a supporting quote from the source material, and our code confirms the quote exists. This week 232 stories went through the review and 13,976 claims were checked. On 1930 claims (13.8%), the reviewers did not all give the same answer. Below are some of those disagreements. We chose them because each involves a real editorial judgment: who said something, what a source actually says, and where reported fact ends and inference begins.

This week in numbers

• 232 stories went through the review (413 review sessions, counting re-checks after corrections); 219 of them are published.
• 13,976 claims checked, 41,542 votes counted (one per reviewer per claim).
• On 1930 claims (13.8%) the reviewers did not all give the same answer.
• Claude (Anthropic): 13,909 votes; backed 96% of the claims it reviewed with a quote we could verify; 0.8% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 417 times.
• GPT-5.6 Sol (OpenAI): 12,963 votes; backed 88% of the claims it reviewed with a quote we could verify; 8.2% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 1132 times.
• Kimi (Moonshot AI): 13,424 votes; backed 96% of the claims it reviewed with a quote we could verify; 2.2% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 261 times.
• Gemini (Google): 1,152 votes; backed 97% of the claims it reviewed with a quote we could verify; 1.2% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 26 times.
• DeepSeek (DeepSeek): 94 votes; backed 67% of the claims it reviewed with a quote we could verify; 23.4% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 4 times.
• Final verdicts: approved 216, sent back for fixes 11, rejected 5.
• In 3 stories at least one reviewer stepped aside because the story concerned the company that built it.

The cases

A contrast the source did not draw

The first draft said that, according to The Guardian, the Labor Department, "rather than Vance," announced Adobe's suspension from the green card labor certification program. Claude (Anthropic) and GPT-5.6 Sol (OpenAI) both rated the claim partial and flagged it. Kimi (Moonshot AI) rated it supported. The recorded note said The Guardian "does not say 'rather than Vance'; that contrast is inferred." The supporting quote was: "The US labor department also said Adobe had been similarly suspended." The question was not whether the Labor Department spoke. It was whether the source drew the contrast the draft implied. The sentence was rewritten to follow The Guardian's wording: the Labor Department also said Adobe was similarly suspended. The corrected draft was re-checked before publication.

Story: US suspends Microsoft, Adobe and others from green card labor certification program

From a weekday to a calendar date

The first draft said Health Cabinet Secretary Aden Duale announced Kenya's first Ebola case on Tuesday, 6 October 2026, according to Reuters, AP, The Guardian and AllAfrica. Claude (Anthropic) and Kimi (Moonshot AI) rated the claim partial and flagged it. GPT-5.6 Sol (OpenAI) rated it supported. The flag note read: "Reuters and Guardian texts do not state Tuesday; exact date is inferred." Turning a weekday into a full date is a small step, but the sources did not take it. The draft also credited the timing to outlets that had not stated it. The correction attributed the Tuesday announcement only to AP, AllAfrica, SCMP and Medical Xpress/AFP, which state it explicitly. It also removed the exact date and noted that Reuters and The Guardian reported the announcement without a date.

Story: Kenya confirms first Ebola case; patient who travelled from DR Congo has died

When a claim about omissions is wrong

The first draft of the Manchester City story ended by saying the reports reviewed "do not include the commission's findings or describe what sanctions, if any, have been imposed." Two reviewers voted on that round. Kimi (Moonshot AI) rated it supported. Claude (Anthropic) rated it contradicted, with the note: "Reports describe findings; sanctions reported as pending, not unknown." It quoted the material: "An independent commission found that City had arranged 'sham' contracts with a number of their commercial partners to artificially inflate revenues and reduce costs." A sentence about what sources leave out is itself a factual claim, and it can be checked like any other. The final sentence was replaced. It now states that sanctions are pending, citing Alonso via BBC Sport. The draft was re-checked before publication.

Story: Manchester City manager Maresca says he expects club to win appeal over financial breaches

A single dissent that changed the text

In the story on Jeffrey Archer's death, the first draft said the reports "do not clarify when this happened relative to the 1986 allegations," referring to his expulsion from the Conservative party. Claude (Anthropic) and Kimi (Moonshot AI) rated the claim supported. GPT-5.6 Sol (OpenAI) alone rated it contradicted, noting that the Guardian "explicitly ties the expulsion to when allegations broke." Its quote: "When the allegations broke, the leader of the Conservative party at the time, William Hague, expelled Archer from the party." Here the minority view led to the correction. The body now carries the Guardian's statement that the expulsion came when the allegations broke. In the same round, the dek was changed from "days before" to "one day before" publication, with Guardian attribution for the timing.

Story: Jeffrey Archer, novelist and ex-MP jailed for perjury, dies at 86, publisher says

Silence versus a reported absence of response

The first draft said the reports reviewed "do not include any response from Trump, the White House or the US government." On that round, Claude (Anthropic) and Kimi (Moonshot AI) rated the claim supported. Gemini (Google) alone rated it contradicted. Its note said the material shows Euronews "checked for and found no Trump reaction as of yet." The quoted line began: "As of yet, no reaction from the subject of 'The Orange Plague'..." The distinction is narrow but real. Saying a source is silent is different from saying the source looked and reported that no response had come. The correction replaced the claim with Euronews's report that no Trump reaction had occurred as of publication. It kept, as a separate statement, that the reports include no White House or US government response.

Story: Gold-coloured satirical statue of naked Trump goes on display at European Parliament

What we learned

Most of these cases turned on one question: does the source say this, or did we infer it? The inference took different forms: an implied contrast, a calendar date worked out from a weekday, and statements about what reports omit. The quote requirement helps because every reviewer must point to the exact words it relies on, and code checks that those words exist. These cases also show that agreement is not proof. In several of them, a lone reviewer's objection led to the correction. We now pay closer attention to sentences describing what sources do not say, which are easy to write and often wrong. We also check attributions that credit more outlets than actually reported a detail.

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

Traceability
Sources (5): Vocemundi: US suspends Microsoft, Adobe and others from green card labor certification program · Vocemundi: Kenya confirms first Ebola case; patient who travelled from DR Congo has died · Vocemundi: Manchester City manager Maresca says he expects club to win appeal over financial breaches · Vocemundi: Jeffrey Archer, novelist and ex-MP jailed for perjury, dies at 86, publisher says · Vocemundi: Gold-coloured satirical statue of naked Trump goes on display at European Parliament
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: 2026-10-11 15:00 UTC

Produced by the Vocemundi newsroom with AI assistance.