Under the hood

Why you can trust this

This section is currently available in English only.

Every summary on this site is generated by a language model. Language models are fluent. They produce confident, well-structured text. They can also make things up ^[1]. Current estimates put hallucination rates at 3–27% depending on the model and task ^[2]. We can't eliminate that entirely. But we can make the credibility mechanisms overt: visible, inspectable, and verifiable by you. Not "trust us." See for yourself.

AI transparency

All summaries on ovr.news are generated by AI using open-source models (Gemma 4 12B for most new articles, Gemma 3 27B for the rest) running on local hardware. Translation is currently switched off. The site runs English-first. Each article shows an AI summary label. AI can make mistakes. That's why we always link to the original article, so you can verify.

Layer 1: Clean the input

A model can only be as faithful as its input. Before any article reaches the language model, it passes a content quality gate. NexusMind runs that gate upstream; when it sends no verdict, our own check runs, with density-based heuristics that look for:

• CSS and HTML leakage: if more than 3–5% of the text is markup, the extraction failed. Rejected.
• Cookie banners: short articles dominated by consent language. Rejected.
• Paywall stubs: articles under 800 characters with phrases like "subscribe to continue." Rejected.
• Navigation debris: when most lines are shorter than 30 characters, you're reading a menu, not an article. Rejected.

Articles that fail the quality gate never reach the language model. Better to show nothing than a confident summary of garbage.

Layer 2: Ground the model

The most effective defense against hallucination is also the simplest: constrain what the model may say. Every summary prompt includes explicit grounding instructions:

Use facts from the article. Do not invent statistics, quotes, or claims. Do not add context, background, or interpretation beyond what the source article states.

Research on retrieval-augmented generation shows that anchoring responses in source documents substantially reduces hallucination rates ^[3]. Our prompts also ban words that signal editorializing rather than reporting: "groundbreaking," "innovative," "significant," "highlights," "showcases," "underscores."

Layer 3: Control the temperature

Language models have a parameter called temperature that controls randomness. Higher values produce more creative, more drift-prone output. Lower values keep the model closer to its input. A 2026 study across 172 billion tokens found that hallucination rates increase measurably with temperature ^[4]. Our summaries run at 0.7, a middle setting, not a low one: constrained enough to stay close to the source material while allowing natural phrasing. The grounding work is done mainly by the prompt constraints and render-time checks.

Layers 1 through 3 are about prevention. Layers 4 and 5 let you check the result yourself; Layer 6 checks it for you.

Layer 4: Link to the source

Every article on this site has a "Read Original" link. The original URL, the source name, and the publication domain are preserved from the moment an article enters the pipeline to the moment it appears on your screen.

Research on AI transparency shows that source attribution is one of the strongest predictors of user trust ^[5]. When other outlets appear to cover the same story, we show how many, but not which: our grouping of stories is not yet precise enough to name an outlet as having reported this one.

Layer 5: Show the scores

Every article shows its weighted average score. Open the article, and you can see which dimensions were scored and how. The filter definitions are not public at the moment; two earlier-generation filters are available on Hugging Face; the five in use today are not public.

You might disagree with a score. That's fine. The point is that the judgment is inspectable.

Layer 6: Editorial rules

After scoring and summarization, an editorial layer makes final decisions. Its rules are deterministic TypeScript checks with configurable thresholds, on top of the pipeline's duplicate removal. Each removal by these rules is logged with a reason; the duplicate fold is only counted.

We also tried an AI rule: an editorial gate in which the model that writes our summaries classified each article (a delivered outcome, a press release, an obituary, off-topic). When we measured it, it rejected about half of all articles while obituaries are about one in twenty, so we switched it off. Obituaries are now caught two ways: a trained classifier upstream reads the source, and a rule checks whether our own summary announces a death.

What we're honest about

• We can't verify facts. The model summarizes what the article says. If the article contains an error, the summary will too.
• Summaries can still drift. Despite grounding prompts and a controlled temperature, subtle distortion happens. A nuance gets lost. An emphasis shifts. This is why the source link exists ^[6].
• Quality depends on extraction. Some websites make it hard to extract clean text. The quality gate usually catches poor extraction. Sometimes it doesn't.
• Scores reflect our lens definitions. The scoring system encodes what we think "belonging" or "discovery" means. We summarize them on the lenses page, but they are still editorial choices.

What AI doesn't do

There are things we deliberately don't leave to AI:

• Whether numbers in an article are accurate
• Whether a source is reliable
• Whether a claim is proven

We leave that judgment to you. We give you the context to decide for yourself.

Source quality

Not all news sources are equal. We assess each source for reliability, so you know where the news comes from.

The tiers

Credibility score

Verified sources, and sources we rate ourselves, get a whole-number score from 0 to 10. Article pages show its band and who gave the rating. A rating describes a publisher's general record. It is never a verdict on one article or one claim. Articles collected before 22 October 2026 show no band.

Where do the scores come from?

External scores come from three databases. They are not three independent opinions:

• IDIAP Research Institute: Academic database with NewsGuard scores and reliability labels for ~5,300 domains. For most of the sources it rates for us, its label relays Media Bias/Fact Check
• Media Bias/Fact Check: Independent assessment of factual reporting and political bias for ~4,400 domains
• Wikipedia Perennial Sources: Community-consensus reliability ratings maintained by Wikipedia editors for ~420 domains

So most externally rated sources rest on the work of one organisation.

For sources not covered by these databases, we assign a score ourselves. Most of these scores were drafted by AI agents and were not checked one by one by a person, so read them as our estimate, not an independent rating. Where our score overlaps with an external database, automated checks flag large disagreements.

Current coverage

We currently track ~1,000 source domains, reached through roughly 1,400 RSS feeds; some domains provide several. Pages elsewhere say "over a thousand sources", which is this number, not the feed count:

Source type

Beyond reliability, we also classify sources by type:

• Wire service: Reuters, AP, AFP
• Academic: Nature, The Lancet, arXiv
• Public broadcaster: BBC, NOS, NPO
• NGO / non-profit: Positive News, Solutions Journalism Network
• Newspaper: The Guardian, El País, de Volkskrant
• Government / institutional: EU, WHO, NIH

Our editorial stance

• We show all tiers. We don't hide articles from curated or unknown sources.
• No score doesn't mean unreliable. It means we couldn't verify.
• A good source can publish a bad article. We assess at the domain level, not per article.

Last updated: July 2026