Last updated: August 15, 2026
Horizon never labels a story true or false. Instead, every story reports the shape of the evidence: who reported it, how independent those sources are from each other, and whether a claim ships with supporting evidence or rests on the claimant's word. You judge; we show the receipts.
Every item is attributed to a publisher — the host its link actually points at, so a Hacker News submission of a TechCrunch article counts as TechCrunch, not Hacker News. Each publisher sits in a tier: primary (first-party and official — the source is the subject: openai.com, arxiv.org, huggingface.co), expert (independent named analysts — simonwillison.net, semianalysis.com), media (news outlets — techcrunch.com, theverge.com), and community (social platforms and aggregators — Hacker News, Reddit, Bluesky — which signal reception, not corroboration).
Near-duplicate stories are clustered, then collapsed to distinct publishers— ten copies of one outlet's article count as one source, and platform aliases (old.reddit.com and reddit.com, x.com and twitter.com) count as one platform. The result is a level: primary source (a first-party publisher reported it), corroborated (two or more independent credible publishers), reported (exactly one credible publisher), or unverified (community discussion with no independent reporting — usually an ordinary conversation, not a claim spreading).
Every curated feed is annotated with where its publisher is based— the newsroom or imprint location from the outlet's own about page. When a story's sources span two or more world regions, the corroboration detail says so: "press coverage spans North America, Europe". This measures publisher location, not independence — wire stories syndicate across borders, so geographic spread is never presented as "independently confirmed in N countries". Where the outlet is based is also a different question from what it covers: a US-based analyst writing about Chinese labs counts as US-based press in the China coverage lens, and both facts are kept.
Countries in East Asia are displayed at region granularity ("East Asia") rather than per country. Community platforms — Hacker News, Reddit, Bluesky, X — are venues rather than publishers and carry no location at all; items that arrive only through them are excluded from the spread rather than guessed.
When a story's central claim is notable — benchmark results, capability assertions, business figures — Horizon records who originated it. A claim is subject-originatedwhen it comes from the entity it benefits (a lab's own benchmark numbers, even when a news outlet is relaying them) and independentwhen it comes from an unaffiliated party (a third-party evaluation, a reproduction, a journalist's own testing).
Subject-originated claims carry an evidence state on the briefing: Vendor-claimed (self-reported, no paper linked, no independent evaluation among the story's sources), Paper linked (the claimant links a paper, technical report, or model card), or Independently evaluated (an independent evaluation appears among the story's sources). "Vendor-claimed" is a fact about sourcing — it says nothing about whether the claim is right.
Each claim also records which entity it concerns, so an entity page can show how the claims Horizon holds provenance for divide up: how many are the entity reporting on itself versus independently evaluated by someone else, over the last 30 days. The denominator is always the claims we hold provenance data for — never everything published about that entity.
Two deliberate limits. Nothing appears at all below three claims, and the plain-language summary ("mostly self-reported") waits for six claims and a clear majority. Counts carry their own sample size in plain sight; a phrase like "mostly" over two items would just name whichever two we happened to see. If an entity shows no provenance block, that means we hold too little to say anything — not that something is wrong with it.
A story upgrades to independently evaluated when its sources include third-party evaluators — currently lmarena, Epoch AI, Artificial Analysis, and METR — or an item whose own claim originates from independent testing. Horizon also ingests Epoch AI and METR directly, so evaluation-methodology work (benchmark saturation, eval construction) surfaces alongside launch coverage.
No truth verdicts. No "unverified" warning badges on ordinary discussion — measured against the live corpus, that label would fire on more than a quarter of items, nearly all of them normal conversation rather than unreported claims spreading, so badging it would be noise. And no sentiment scoring of people. Every signal on this page is also available to agents via the API and MCP server as structured fields.
See the signals live on today's briefing or start at the front page.