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: 11 October 2026 to 18 October 2026 (UTC).

Each week, three AI reviewers, Claude (Anthropic), GPT-5.6 Sol (OpenAI) and Kimi (Moonshot AI), check every claim in our stories against the source material. Each must copy a supporting quote, and our code confirms that the quote exists. A claim needs two reviewers backing it with a verified quote. This week 220 stories went through review and 14,138 claims were checked. On 2360 claims the reviewers did not all give the same answer. Below are five of those cases, chosen because each turns on a real editorial judgment.

This week in numbers

• 220 stories went through the review (423 review sessions, counting re-checks after corrections); 203 of them are published.
• 14,138 claims checked, 41,909 votes counted (one per reviewer per claim).
• On 2360 claims (16.7%) the reviewers did not all give the same answer.
• Claude (Anthropic): 14,029 votes; backed 95% 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 539 times.
• GPT-5.6 Sol (OpenAI): 13,526 votes; backed 88% of the claims it reviewed with a quote we could verify; 8.6% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 1140 times.
• Kimi (Moonshot AI): 11,336 votes; backed 94% of the claims it reviewed with a quote we could verify; 4.7% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 516 times.
• Gemini (Google): 2,821 votes; backed 98% 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 80 times.
• DeepSeek (DeepSeek): 197 votes; backed 80% of the claims it reviewed with a quote we could verify; 16.2% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 7 times.
• Final verdicts: approved 194, sent back for fixes 17, rejected 9.
• In 4 stories at least one reviewer stepped aside because the story concerned the company that built it.

The cases

Does the source actually name the city?

A Science story on the Magnitude 6.9 earthquake on Russia's Kamchatka Peninsula stated that Petropavlovsk-Kamchatsky is the region's administrative centre. GPT-5.6 Sol (OpenAI) and Kimi (Moonshot AI) marked it supported. Both cited a line from The Hindu: "A number of residents of the region's administrative centre evacuated from their apartments because of tremors, it said." Claude (Anthropic) marked it unsupported, noting that the material "mentions an 'administrative centre' but never names Petropavlovsk-Kamchatsky as it." The split turns on a narrow question: does a sentence that refers to the administrative centre also establish which city that is? The claim was published as written, because at least two reviewers backed it with a verified quote.

Story: Magnitude 6.9 earthquake strikes Russia's Kamchatka Peninsula

A lone objection that changed the draft

A Culture story on the fire at Paris's oldest covered food market said in its first draft that the reports "do not describe when the fire started or how many firefighters were deployed." In that round the reviewers were Claude (Anthropic), GPT-5.6 Sol (OpenAI) and Gemini (Google). Claude and Gemini marked the sentence supported. GPT-5.6 Sol marked it contradicted, noting that the reports "give an approximate start period, though not firefighter deployment numbers." It pointed to the line: "The historic Marché des Enfants-Rouges in central Paris was gutted by fire early on Monday." The draft was corrected in the next review round. It now says that France 24 and Reuters report the fire began early Monday, without giving an exact time. It keeps the statement that the reports do not say how many firefighters were deployed. The story was then re-checked before publication.

Story: Fire destroys Paris's oldest covered food market overnight, authorities say

Who made the privacy appeal?

A Health story on Kenya's second Ebola case said in its first draft that the ministry urged the public to "stay calm but vigilant", according to the BBC, "and to respect the family's privacy." Claude (Anthropic) and GPT-5.6 Sol (OpenAI) marked it partial. Kimi (Moonshot AI) marked it supported. The flag note explained: "Privacy appeal is from AllAfrica, passive voice; not BBC or explicitly ministry." In the next round the privacy appeal was separated from the BBC quote and attributed to AllAfrica, in passive voice as reported. A following sentence was changed to "the ministry says" for clarity. In the analysis, "one household" became "close relative", matching the reports. The draft was then re-checked before publication.

Story: Kenya confirms second Ebola case in close relative of first patient

Matching texts, but a shared source?

A Climate story on a report about falling global forest loss said in its first draft that AFP, as carried by Phys.org, and the South China Morning Post "use identical wording and appear to come from the same AFP dispatch." All three reviewers marked the sentence partial that round. The flag came from GPT-5.6 Sol (OpenAI) and Kimi (Moonshot AI), not from Claude (Anthropic). The flag note read: "The texts closely match and Phys.org credits AFP; shared-dispatch origin is not explicitly stated." In the next round the claim was replaced with a statement that the reports' texts closely match and that Phys.org credits AFP. A second change replaced "findings were presented Wednesday" with the reports' own wording: the study came out Wednesday and a co-author spoke at a briefing. The draft was then re-checked before publication.

Story: Report says global forest loss fell 15% in 2025 as Congo losses neared Amazon's

Where a reported fact becomes a comparison

A Sports story on LIV Golf's future included an analysis line: if LIV 2.0 proceeds, its player-majority ownership model "would be an unusual arrangement in professional sport." Kimi (Moonshot AI) marked it supported, citing ESPN's reported equity split among golfers, BC Partners and potential minority investors, and management. Claude (Anthropic) cited the same passage but marked it partial, noting: "'Unusual in professional sport' asserted without hedge; not grounded in material." GPT-5.6 Sol (OpenAI) marked it unsupported, writing that the sources "give ownership percentages but do not establish that such an arrangement is unusual." The reviewers agreed on the ownership facts and differed on the comparison drawn from them. The claim was published as written, because at least two reviewers backed it with a verified quote.

Story: LIV Golf decision on wind-down or 'LIV 2.0' expected by end of October

What we learned

In these cases the reviewers rarely disagreed about the facts themselves. They disagreed about what a source does and does not say. Some flags separated an inference from a reported statement, such as a shared dispatch or an "unusual" arrangement. Others caught attribution that drifted to the wrong outlet or speaker. A missing detail differs from an approximate one, as the Paris fire showed. A sentence that mentions an administrative centre may not name it. Requiring a verified quote makes every vote checkable. Under our two-quote rule, a quote attached to a partial vote also counts as backing, as in the LIV Golf case. Where drafts changed, the fixes were modest: hedging, re-attribution, and replacing assertions with the sources' own wording.

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

Traceability
Sources (5): Vocemundi: LIV Golf decision on wind-down or 'LIV 2.0' expected by end of October · Vocemundi: Report says global forest loss fell 15% in 2025 as Congo losses neared Amazon's · Vocemundi: Fire destroys Paris's oldest covered food market overnight, authorities say · Vocemundi: Kenya confirms second Ebola case in close relative of first patient · Vocemundi: Magnitude 6.9 earthquake strikes Russia's Kamchatka Peninsula
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: 2026-10-18 15:01 UTC

Produced by the Vocemundi newsroom with AI assistance.