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: 18 October 2026 to 25 October 2026 (UTC).

Every Vocemundi story is broken into individual claims. Three AI reviewers, Claude (Anthropic), GPT-5.6 Sol (OpenAI) and Kimi (Moonshot AI), mark each claim supported, partial, unsupported or contradicted. Each reviewer must copy a supporting quote from the source material, and our code checks that the quote exists. A claim needs two reviewers backing it with a verified quote. This week 225 stories went through review and 13,372 claims were checked. On 2370 claims the reviewers did not all give the same answer. Here are five of those splits and how they were settled.

This week in numbers

• 225 stories went through the review (404 review sessions, counting re-checks after corrections); 214 of them are published.
• 13,372 claims checked, 39,777 votes counted (one per reviewer per claim).
• On 2370 claims (17.7%) the reviewers did not all give the same answer.
• Claude (Anthropic): 13,312 votes; backed 95% of the claims it reviewed with a quote we could verify; 1.4% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 506 times.
• GPT-5.6 Sol (OpenAI): 12,698 votes; backed 87% of the claims it reviewed with a quote we could verify; 8.9% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 1107 times.
• Kimi (Moonshot AI): 13,332 votes; backed 93% of the claims it reviewed with a quote we could verify; 5.5% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 655 times.
• Gemini (Google): 366 votes; backed 99% of the claims it reviewed with a quote we could verify; 0.3% 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): 69 votes; backed 97% 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 1 times.
• Final verdicts: approved 203, sent back for fixes 15, rejected 7.
• In 6 stories at least one reviewer stepped aside because the story concerned the company that built it.

The cases

Matching text versus a stated credit

The claim said The Hindu also published the Reuters report on police using tear gas against student-led protesters in Tirana. Kimi (Moonshot AI) marked it supported and quoted The Hindu's headline. Claude (Anthropic) rated it partial. Its note said the Hindu text matches Reuters, "but the material shows no explicit Reuters credit." GPT-5.6 Sol (OpenAI) marked it unsupported and noted that the material "does not explicitly say The Hindu republished Reuters." This is a familiar question in news checking. Matching text can point to a shared source, but an explicit credit is what states it. Under our rule, the claim was published as written because at least two reviewers backed it with a verified quote. The split is still a useful reminder of the gap between resemblance and stated attribution.

Story: Albanian police use tear gas on student-led protesters in Tirana, Reuters reports

Which outlet supplied which details

A first draft about the six-year ban on Ethiopian marathon runner Ruti Aga said that all details beyond the ban came from The Guardian alone. It also said the Reuters item was a summary only. Claude (Anthropic) rated this partial and Kimi (Moonshot AI) marked it supported. Neither flagged it. GPT-5.6 Sol (OpenAI) marked it contradicted and flagged it, noting that Reuters "also supplies the substance, tampering characterization, and 2019 Tokyo win." Its supporting quote named the Tokyo win, the substance methylprednisolone and tampering with evidence. The draft was corrected in the next round. The sourcing line was replaced, and both outlets were described as attributing their reports to the AIU. A statement that the AIU treated the prescription as tampering was removed, with a note that The Guardian does not explicitly link them. The story was re-checked before publication.

Story: Ethiopian marathon runner Ruti Aga banned six years over prohibited substance and tampering

A likely data source is not a named one

A first draft on Vance strengthening into a Category 1 hurricane said both news agencies "likely rely" on data from the U.S. National Hurricane Center, "but only the AP says so." Claude (Anthropic) marked it supported. GPT-5.6 Sol (OpenAI) and Kimi (Moonshot AI) rated it partial and flagged it. The recorded note reads: "AP names NHC; Anadolu material does not support likely reliance." The supporting quote was: "The National Hurricane Center reports that Vance has become a hurricane off the coast of Mexico." In the next round, the speculation was replaced with what each agency actually states. Other edits removed an unsupported editorial confidence rating sentence. They also changed "maximum winds" to "winds" to match AP wording and clarified which details came from which agency. The draft was re-checked before publication.

Story: Vance strengthens into Category 1 hurricane off Mexico's coast

Two reported facts, one unstated link

A first draft about the publisher Boréal ending ties with author Thélyson Orélien said Boréal's conclusion rested on two things. One was metadata findings by an unnamed analyst. The other was an earlier detector test whose error rate the reports did not describe. Claude (Anthropic) and GPT-5.6 Sol (OpenAI) rated it partial and flagged it. Kimi (Moonshot AI) marked it supported. The flag note said the material "doesn't say Boréal's conclusion relied on the Pangram test." The supporting quote connected that test to something else: a withdrawal "from the shortlist for France’s top book prize, the Goncourt." In the corrected analysis, Boréal's conclusion is attributed only to the metadata findings. The Pangram test is tied to the shortlist withdrawal and noted as not described as part of Boréal's decision. The draft was then re-checked before publication.

Story: Publisher Boréal ends ties with author Thélyson Orélien, says AI was used in novel

Listing an outlet for a finding it did not report

The first draft's lead said a UK public inquiry found that government failures contributed to the spread of the virus in care homes, "as reported by BBC News, The Guardian and RT." GPT-5.6 Sol (OpenAI) marked it supported. Claude (Anthropic) and Kimi (Moonshot AI) rated it partial and flagged it. The flag note said the RT headline "concerns blame for deaths, not virus spread specifically." The quoted headline read: "Inquiry blames UK government oversight for high number of Covid-19 deaths in care homes during pandemic." The difference is narrow but real: naming an outlet next to a finding implies it reported that specific finding. The lead was corrected to attribute the virus-spread finding to BBC News and The Guardian only. It now adds that RT reported the inquiry blamed government oversight for high care home deaths. The draft was then re-checked.

Story: UK Covid inquiry finds government failures contributed to virus spread in care homes

What we learned

In these cases the facts were largely in the material, but the wording around them said more than the sources did. Matching text is not a stated credit. A likely data source is not a named one. Two facts reported near each other do not establish that one caused the other. A headline about deaths is not a finding about spread. In the cases above, each reviewer at some point flagged something the others accepted. This is why we require verified quotes and two reviewers backing each claim. It is also why flagged drafts go back for another round and are re-checked before publication. The corrections this week were mostly about attribution: who said something, and exactly what they said. That is what we watch most closely.

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

Traceability
Sources (5): Vocemundi: Albanian police use tear gas on student-led protesters in Tirana, Reuters reports · Vocemundi: Ethiopian marathon runner Ruti Aga banned six years over prohibited substance and tampering · Vocemundi: Vance strengthens into Category 1 hurricane off Mexico's coast · Vocemundi: Publisher Boréal ends ties with author Thélyson Orélien, says AI was used in novel · Vocemundi: UK Covid inquiry finds government failures contributed to virus spread in care homes
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: 2026-10-25 15:01 UTC

Produced by the Vocemundi newsroom with AI assistance.