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 October 2026 to 4 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 code checks that the quote exists. This week 5,324 claims were checked, and on 781 claims the reviewers did not all give the same answer. Below are four of those cases, chosen because they turn on what a source actually says. For each, we set out the claim, the reviewers' recorded positions and the outcome.

This week in numbers

• 89 stories went through the review (157 review sessions, counting re-checks after corrections); 80 of them are published.
• 5,324 claims checked, 15,943 votes counted (one per reviewer per claim).
• On 781 claims (14.7%) the reviewers did not all give the same answer.
• Claude (Anthropic): 5,324 votes; backed 97% 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 106 times.
• GPT-5.6 Sol (OpenAI): 5,162 votes; backed 88% of the claims it reviewed with a quote we could verify; 8.1% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 504 times.
• Kimi (Moonshot AI): 5,174 votes; backed 96% of the claims it reviewed with a quote we could verify; 1.9% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 141 times.
• Gemini (Google): 213 votes; backed 92% of the claims it reviewed with a quote we could verify; 6.6% of its votes came without a quote that our code could find in the source material; it stood alone against the other two 13 times.
• DeepSeek (DeepSeek): 70 votes; backed 97% 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 0 times.
• Final verdicts: approved 84, sent back for fixes 5.
• In 2 stories at least one reviewer stepped aside because the story concerned the company that built it.

The cases

Does a possible grand jury mean no charges?

The story on the alleged Cornell rape case said: "No charges have been filed, according to the reports reviewed by Vocemundi." Claude (Anthropic) marked it supported, citing an Al Jazeera line: "If warranted, evidence will be presented to a grand jury and the case prosecuted." GPT-5.6 Sol (OpenAI) marked it partial, noting: "Sources confirm no 2024 charges, but not explicitly no current charges." Kimi (Moonshot AI) marked it unsupported and gave no quote, noting that the reports state the 2024 closure without charges but not the charge status after the case was reopened. Claude was alone in marking the claim fully supported. Resolution: published as written, because at least two reviewers backed the claim with a verified quote.

Story: New York governor names attorney general special prosecutor in alleged Cornell rape case

A predicted busy weekend, reported as fact

The first draft's summary line said Munich's Oktoberfest passed its midpoint "after a busy 'Italian weekend'." Claude (Anthropic) and Kimi (Moonshot AI) marked the claim partial. GPT-5.6 Sol (OpenAI) marked it supported. The flag note reads: "Source predicts a traditionally busy weekend; it does not report this one was busy." The quoted source material says the "traditionally particularly busy" weekend "is just around the corner". The wording was changed to "around the traditionally busy Italian weekend". In the same round, the body separated assault, harassment and upskirting reports from the offence total, matching DW's "also filed" wording. The analysis now attributes the overcrowding-prevention aim to Euronews and limits it to cameras and AI tracking. The draft was corrected in the next review round and then re-checked before publication.

Story: Munich's Oktoberfest heads into its final weekend after an estimated 3.8 million visitors by midpoint

Unnamed officials are not necessarily the same officials

The first draft's analysis of the SpaceX Crew-13 story said the record claim "rests on NASA's account, since both outlets cite the same unnamed officials". Claude (Anthropic) and GPT-5.6 Sol (OpenAI) marked it partial. Kimi (Moonshot AI) marked it supported. The flag note reads: "Material doesn't show both outlets cite the same officials; asserted as fact." The quoted source text attributes the record to "NASA officials" without naming them. The unsupported assertion was removed, and the analysis now says both outlets attribute the record to unnamed NASA officials. The same round also corrected a body claim about DW's reporting: the crew's arrival, not the launch, was planned for mid-September before a leak delay. The draft was corrected in the next review round and then re-checked before publication.

Story: SpaceX Crew-13 reaches ISS in under eight hours, a US record, NASA officials say

A claim of no police comment, contradicted

The first draft of the story on 49ers owner Jed York said: "The reports include no comment from prosecutors, police or York's legal representatives." GPT-5.6 Sol (OpenAI) and Kimi (Moonshot AI) both marked this contradicted. Claude (Anthropic) marked it partial and is not recorded among the reviewers who flagged it. The flag note says The Guardian includes a police-attributed statement about York's alleged conduct, quoting: "police said he was attempting to pay $160 for sex with a woman he believed to be a prostitute." The claim about police was removed, and the statement was kept for prosecutors and York's legal representatives. A sentence was added noting that the only police account is the one cited by The Guardian on the alleged payment. The draft was corrected in the next review round and then re-checked before publication.

Story: NFL suspends 49ers owner Jed York six games, fines him $500,000

What we learned

In each of these cases the disputed wording went a step beyond the source. A forecast was written as a report. Two separate attributions were written as one shared source. A statement that the reports contained no comment overlooked a police-attributed line. A sentence about a possible future grand jury was read as confirming the present status of charges. Requiring a copied quote makes such gaps visible, but reviewers still differ on how much a given quote proves. Three of the drafts were corrected and re-checked. The fourth claim was published as written under the two-reviewer rule, although one reviewer marked it unsupported. We watch for claims about what reports do not contain, for tense, and for statements about where reporting originated.

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

Traceability
Sources (4): Vocemundi: New York governor names attorney general special prosecutor in alleged Cornell rape case · Vocemundi: Munich's Oktoberfest heads into its final weekend after an estimated 3.8 million visitors by midpoint · Vocemundi: SpaceX Crew-13 reaches ISS in under eight hours, a US record, NASA officials say · Vocemundi: NFL suspends 49ers owner Jed York six games, fines him $500,000
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: 2026-10-04 15:00 UTC

Produced by the Vocemundi newsroom with AI assistance.