ClaimTrace FAQ
Honest answers — including about what we do not do.
What is ClaimTrace?
ClaimTrace is a verification layer built by AXIONYX.io. Its AI opens the entity's own sources — website, linked PDF, scanned image — and returns the verified fact as structured JSON with its proof: source URL, date, verbatim excerpt and a confidence score. When nothing is published, it returns a motivated null instead of a plausible guess.
How is ClaimTrace different from a search or SERP API?
A search API tells you what a search engine displays. ClaimTrace tells you what the entity itself publishes: it reads the entity's own pages and documents and hands you the verbatim as proof. Aggregators are explicitly rejected as sources — 17 of them — because a fact you cannot trace to its owner is unverifiable by construction.
What proof comes with each fact?
Every claim carries its provenance: the source URL, the date it was read (ISO 8601), a short verbatim excerpt from the source, the kind of page it came from, whether the domain belongs to the entity, a confidence score between 0 and 1, and a live-verification tri-state (confirmed, contradicted or unverifiable).
What happens when the data simply is not published?
You get an explicit not_found with the reason, a detail, and the source that was actually checked. Absence is a documented answer — it is never filled in with something plausible. That is exactly what an AI agent can handle without hallucinating.
Can ClaimTrace hallucinate data?
Not by construction: a served fact must exist literally in what was read, and a truncated document is repaired without a single word added. Across the last four external adversarial audits, the auditors found no fabricated menu — the last two alone covered 91 entities checked one by one (31 July 2026). The engine has been through 11 external adversarial audits, and 3 299 automated tests ran green on the full suite re-run of 8 August 2026.
Which sources does ClaimTrace accept?
Only the entity's own publications, enforced by a domain-ownership whitelist. An anti-homonym identity gate makes the address outweigh the name, so a business is never confused with a namesake.
Can it read JavaScript pages, PDFs and scanned images?
Yes. Pages are rendered in a real browser, linked PDFs are opened, and scanned images are read — the fact comes from the real content, not from a cached snippet.
How fresh is the data?
Before a cached fact is served, it is revalidated live and marked confirmed, contradicted or unverifiable — and reported as such. Dated data is cached for at most 24 hours, and cache hits are served free of charge.
How much does it cost?
One credit is one entity freshly resolved and verified. Beta launch pricing: Discovery is free (50 credits plus 25 per month), Pack S is 15 CHF for 150 credits, Pack M is 45 CHF for 550, Pack L is 140 CHF for 2000, and Scale is 99 CHF per month with your own AI model key. Payment in CHF via Stripe. Refusals (rate limit, exhausted credits, invalid request) are never charged. This grid is indicative for the private beta and may be adjusted at commercial launch.
Is ClaimTrace production-proven?
ClaimTrace is not a demo: it is the production engine behind arc-menu.ch. 20 out of 20 customer-journey assertions (key, quota, credits, isolation) were verified live on the production API on 31 July 2026, and 9 anchored golden cases are re-verified against the real sources before every deployment — one deviation blocks the release.
What does ClaimTrace not do?
It does not resell Google Places content: ratings, review counts and price levels derived from Places are excluded from the payload for legal reasons. It does not rank pages — it establishes facts. And the restaurant vertical is currently the most mature; other verticals run in generic beta. When we do not know, the API says so.
How do I get access?
Signup is self-serve: send your email to the signup endpoint and receive an API key (shown only once) with 50 trial credits. The beta opens client by client — for anything else, write to kam@axionyx.io.