Harbi: news rebuilt around the event, for readers who are done with the feed

The Reuters Institute’s Digital News Report — the largest ongoing survey of news audiences in the world — has spent years documenting a quiet exodus: around four in ten people now say they often or sometimes avoid the news, and trust in it has been stuck near 40%. The instinct is to blame the audience. We think the audience is behaving rationally, because the feed it is avoiding really did get worse. As we wrote in August, NewsGuard was counting 3,749 AI-generated “news” sites by late June — growing by hundreds a month — and Graphite’s crawl analysis found machine-written articles crossed half of new output back in 2024. Somewhere under that flood there is still the one honest question a reader opens a news site to answer — what really happened? — and the product shape that dominates news on the internet is no longer built to answer it.

The article was the unit. The feed made it the problem

Follow one development through a normal news day and the shape of the problem is hard to unsee. An agency files a wire story; within the hour, dozens of outlets publish the same text under different headlines. Each rewrite is a new article, each article a new click, each click an argument for another notification. Volume masquerades as verification: the reader sees “everyone is reporting this” when, at the root, one newsroom reported it and the rest copied. Headlines are rewritten silently, so what a story said at 09:00 and what it says at 14:00 are two different claims with no record between them. And the feed never ends — not because the day doesn’t, but because the feed is not optimised for the question. It is optimised for the time you spend failing to answer it.

Change the unit: the event

Harbi is built on a different atom. The unit is not the article but the event: dozens of stories about the same development merge into one page. The page’s address is permanent; as the story develops, the page is versioned — updated in the open, with the whole history visible — instead of being silently rewritten. And sources are counted the way an analyst would count them: outlets carrying the same agency wire collapse into one root, so “34 outlets” is displayed as what it actually is — say, six genuinely independent sources — with official statements marked as official rather than laundered into neutral confirmation. Thirty-four echoes are not thirty-four confirmations, and a news product should say so on the page.

Evidence per sentence — and “we don’t know yet” out loud

The second discipline runs deeper than the page: it runs through the prose. Harbi’s design rule is that every published factual sentence is bound to the sources it rests on — names, numbers, dates, causal claims and quotes do not ship without evidence, and what cannot be evidenced either stays out or is filed, explicitly, under “we don’t know yet”. Before anything publishes, automated quality gates check source coverage, number and date consistency, risky language and personal data; content that fails the threshold does not go out. After publication the honesty continues: a correction ships as a new visible version with the old one archived, a fully wrong event is retracted with its page left standing and marked as retracted, and every correction is listed on a public corrections page. The audio layer keeps the same rule — hourly summaries and scheduled bulletins are read by a synthetic voice and labelled as exactly that, the norm Article 50 of the EU’s AI Act has made law in Europe since August 2025, and the one we argued for in our post on the voice-disclosure era.

Where Harbi stands

Harbi is our AI-native news product, live in open beta at harbi.me — in Turkish, for the Turkish news cycle, one of the fastest and noisiest in the world, which is precisely why we built the discipline here first. The day page opens with a question no feed asks: how much time do you have? Two, five or fifteen minutes produce the same day at different depths — not different days. A running course-of-the-day timeline carries developments with their trust math attached — outlets, independent roots, versions — while hourly audio summaries and scheduled bulletins cover the commute, local editions cover the city, and an ask-Harbi answer layer sits on the same evidence-bound corpus. Underneath, it is the platform this site describes everywhere else: our own models, through our own gateway.

• Event pages, not articles: one permanent, versioned page per development
• Sentence-level evidence — every factual claim is bound to its sources
• Independent-source counting: agency copies collapse into one root
• Automated pre-publication quality gates; visible corrections and retractions
• A day with an end: digests at 2 / 5 / 15 minutes, hourly audio bulletins with a labelled synthetic voice

What we will not claim

Three honest limits. First, automated news can be wrong, and we will not pretend otherwise. The promise is not infallibility — it is that every factual sentence shows what it stands on so you can check us, and that being wrong is expensive for us in public: versioned, listed, archived, never silently patched. Second, Harbi is Turkish-language and Türkiye-focused today; we would rather point at one market where the discipline demonstrably works than gesture at a global product that does not exist yet. Third, the absence of a feed is a trade, and we state it: Harbi does not track people, does not tailor the news to your profile, and ends the day on purpose — it measures whether you are informed, not how long you scroll. If what you actually need is always-on monitoring of the public web at organisational scale, that is a different product, and we build that too: Pickerman.

Thirty-four outlets are not thirty-four confirmations. Rebuild news around the event, bind every sentence to its evidence — and say “we don’t know yet” out loud.

For readers in Türkiye the interesting part may be the architecture as much as the beta. The disciplines under Harbi — citation-first generation, independence counting, visible versioning, labelled synthetic audio — are the same ones this platform applies elsewhere: Solab’s notes cite the exact moment a claim was said, Pickerman synthesises only from three or more independent sources, Blab’s voices ship with disclosure built in. Harbi is what happens when that stack is pointed at the hardest consumer surface there is: the daily news — with personal data screened in the quality gates, the way a product born under the KVKK should be built. It is in open beta and reading is free — harbi.me, on our own models through our own gateway, from Arpanet Bilişim A.Ş. Judge it by the check you can run on any page: pick a sentence, and look at what it stands on.