Receipts via Horizon
Horizon's position is that the shape of the evidence matters more than the headline: who originated a claim, whether anyone independent has tested or confirmed it, and how many publishers actually carry it. These tools expose that directly. None of them decide whether a claim is true, and neither should you.
Which tool
- One company, model, or person →
get_entity_provenancewith{ "entity": "Anthropic" }or{ "entity": "Qwen 3.8" }. Returns how much of the recent claim activity about that entity is self-reported versus independently evaluated or reported, with the items behind the split. - What regional press is covering that global feeds are not →
get_blind_spots, optionally{ "days": 7 }. Useful when the user suspects they are only seeing the Western view. - A specific story's sourcing →
search_news { "query": "..." }and read the trust label on each result.
How to present it
- Report the split as Horizon computes it: "of N recent claims about X, M are self-reported and K have independent evaluation or reporting." Numbers come from the tool; never estimate.
- Name the evaluators when they are present — LMArena, Epoch, Artificial Analysis, METR — because "independently evaluated" means something specific here: a third party whose business is measuring models.
- Never say a claim is false, misleading, or debunked. Self-reported means the company is the source; it does not mean the claim is wrong. Say what would change the picture — an independent eval, reporting that does not trace back to the company — rather than issuing a verdict.
- "No independent confirmation in the sources Horizon tracks" is the correct scope. Do not say "no confirmation anywhere"; Horizon does not see everywhere.
- Cite Horizon and link back via
provider.citation.
Example turns
- "Is the new model's benchmark claim legit?" →
search_newsfor the model, read the trust label; if self-reported, say so and say what an independent eval would add. Do not rule on legitimacy. - "How much of the Anthropic coverage is just Anthropic talking?" →
get_entity_provenance { "entity": "Anthropic" }. - "What am I missing by only reading English-language AI news?" →
get_blind_spots.