Ground your AI agent in verified facts

A generative model produces plausible text. When your agent needs the opening hours of a clinic, the daily menu of a restaurant or the services of a local business, "plausible" is precisely the problem: a plausible answer that is wrong reaches your user with your name on it.

ClaimTrace closes that door by construction. It is not another model guessing better — it is a verification layer your agent calls as a tool: you ask for a field on an entity, and you receive either the fact with its proof, or an explicit, motivated absence. Both are things an agent can handle without inventing anything.

What your agent receives

{
  "field": "menu_du_jour",
  "value": "…",
  "provenance": {
    "source_url": "https://…",
    "fetched_at": "2026-07-31T09:10:12Z",
    "verbatim": "…",
    "domain_owned_by_entity": true
  },
  "confidence": 0.93,
  "verified_live": true
}

Every claim carries the source URL, the date it was read, a verbatim excerpt you can display to your user, and a confidence score. When the entity publishes nothing, the response is a not_found with the reason and the source that was checked — your agent can say "not published" instead of making something up.

Why this beats "just add RAG"

Retrieval grounds your model in *documents you already have*. ClaimTrace grounds it in *what the entity itself publishes right now*: pages rendered in a real browser, linked PDFs and scanned images read, aggregators rejected as sources, and a live revalidation before anything is served from cache. The answer is not "similar to something in your index" — it is traced to its owner, with the verbatim as evidence.

Wire it into your pipeline

One blocking POST /v1/fetch for tool calls, one SSE stream when you want entities as they resolve. Stable JSON contract (FetchResultV1), explicit 401/402/422/429 errors, degradations declared in warnings. See the quickstart — you can be in production in three requests.

Proven, not promised

ClaimTrace is the production engine behind arc-menu.ch. It went through 11 external adversarial audits; across the last four, auditors found no fabricated menu — the last two alone covered 91 entities checked one by one (31.07.2026). 3 299 automated tests ran green on the full suite re-run of 08.08.2026. The numbers we publish are the audited ones — nothing else.