How we verify every story before publishing it
When an AI assistant answers news questions unchecked, it gets something significant wrong about 45% of the time (international EBU/BBC study, 2025). Lumina is built specifically not to work that way: reliability doesn't come from models agreeing, but from verification against sources and a person who has the final say.
Produced by an AI pipeline with human editorial oversight.The choices that hold reliability together
Independent verification
A check re-opens every cited source and looks for confirmation elsewhere. Central facts that aren't confirmed block publication.
Human oversight
A person approves the topics and decides on dubious cases. The final word is always human.
Grounded in sources
Every article is written by reading the original pages, not from the model's memory. No claims without a verifiable source.
Source transparency
Every entry carries its links, in plain sight. You can always trace back what we read.
Several models for coverage
Three models from different vendors search in parallel, so an important story is less likely to slip past them all.
Memory across editions
A knowledge base built on semantic similarity links each edition to past ones: it spots what's new, avoids duplicates, and suggests related reading.
From the news stream to the published edition
Three independent models gather, one model synthesises, then a chain of checks decides what deserves to be published. The two highlighted steps are where something can be stopped.
How an edition comes together, step by step
Several models, in parallel
Three frontier models from different vendors (today Claude by Anthropic, Gemini by Google, and GPT by OpenAI) read the day's news with live web search, each on its own. Independent viewpoints serve coverage above all: an important story is less likely to slip past all three.
Synthesis and corroboration
One model compares the three reports: it merges the same story reported by several, tracks how many flagged it, ranks by relevance, and assigns a confidence level. Agreement between models helps prioritise — it is not proof of truth.
Comparison with past editions
Before writing, every topic is compared against the archive of already-published editions using semantic search (local embeddings, no external services). What is genuinely new is told apart from a mere update, duplicates are avoided, and a knowledge base grows over time.
Editorial selection
Based on relevance, confidence, and novelty, a person decides which topics to cover: include, hold, or discard. Selection stays human.
Writing grounded in sources
Each article is written by actually reading the original pages, with mandatory inline citations and at least two or three independent sources. In Italian, in a sober style.
Adversarial verification
A check re-opens every cited source and, on top of that, runs independent searches for each claim. It labels every fact as supported, unsupported, or contradicted. If a central fact doesn't hold, publication is blocked.
Arbiter
If verification fails, an arbiter decides: a targeted revision, or escalation to a person. Nothing is published by guessing.
Text quality
The Italian writing goes through rule-based checking tools — not just language models: Vale for style, LanguageTool for grammar, textstat for readability. The flagged corrections are then applied to the text, leaving facts, sources, and citations untouched.
Translation
The English version is translated and then re-checked, adversarially, for faithfulness to the original.
Transparent publication
The bilingual edition reaches the site with every source linked and links to related past editions, verifiable by anyone.
The limits we know about
No automated system is perfect, and we'd rather say so ourselves. Here are the limits we know about and how we handle them.
Agreement between models is a hint, not proof
Different models can be wrong in the same way, especially on a story they read from the same source. That's why we treat corroboration as a prioritisation signal, never as a test of truth: that's the job of the source check.
The verifier is itself an AI
The automated check can be wrong. We anchor it to external evidence and pair it with human escalation, but we don't publish a measured error rate for it.
Live search, with no fixed list of sources
We don't use a fixed list of approved outlets: coverage stays broad, but the quality of an individual source depends on the models' judgement. Corroboration and the block on unsupported facts are the safety net.
Human oversight is at the key points
People approve topics and read whatever is flagged as dubious, but an article that passes all the automated checks can be published without a line-by-line read.
A daily roundup has blind spots
A story no model finds that day simply doesn't appear, and a fast-moving event can be caught mid-correction. The freshest news is also the hardest to verify.
Spotted an error?
Write to us and we'll fix it. Transparency is worth more than the appearance of perfection.
What this approach is built on
The pipeline's choices follow recognised academic research and editorial standards. A roundup that values sources couldn't fail to cite its own.
- The study measuring how often AI assistants get the news wrong without checks.EBU/BBC — AI assistants misrepresent news 45% of the time (2025)
- The analysis of incorrect citations in AI search engines.Tow Center / Columbia Journalism Review — AI Search Has a Citation Problem (2025)
- The factuality method that breaks text into claims and checks each against independent searches.Long-form factuality in large language models — SAFE (Google DeepMind, 2024)
- The technique for reducing hallucination by verifying claims independently of the draft.Chain-of-Verification Reduces Hallucination in LLMs (2023)
- The research showing why a model cannot correct itself without external evidence.Large Language Models Cannot Self-Correct Reasoning Yet (ICLR 2024)
- The study showing how citations, on their own, often don't actually support what they claim.Evaluating Verifiability in Generative Search Engines (Stanford, 2023)
- The work explaining why blindly mixing several models doesn't pay off — and why we don't do it.Rethinking Mixture-of-Agents — Self-MoA (2025)
- The deontological charter on the use of AI in journalism.Paris Charter on AI and Journalism — RSF
Now that you know how it works
Read the latest edition or get it by email.