The AI that reads the source Private beta

An AI that reads the real sources. And proves it.

READ. ANCHOR. PROVE.

ClaimTrace is a verification layer. Its AI opens the entity's own publications — full web page, linked PDF, scanned image, event pop-up — reads them, and returns the field you asked for with its proof: source URL, date, verbatim excerpt, confidence. Nothing published? A motivated null. Never an invention.

Get a free key Try it in the console

An AI that reads — never guesses 11 external adversarial audits 0 fabricated menu in the last 4 audits

ClaimV1 — real API response · Musée du Louvre, Paris · prod 09.08.2026

{
  "field": "tarif_plein_adulte",
  "value": "22 € (résidents/ressortissants EEE) ou 32 € (hors EEE)",
  "provenance": {
    "source_url": "https://www.louvre.fr/visiter/horaires-tarifs",
    "fetched_at": "2026-08-09T16:47:48Z",
    "verbatim": "Tarif France et EEE … 22.00€ Tarif hors EEE … 32.00€",
    "domain_owned_by_entity": true
  },
  "confidence": 0.97
}

Verbatim extract of a response captured in production on 09.08.2026 (req_b0b08fbddcbf4effb153). The source is the museum's own website, publicly consultable. The AI read that page — it did not summarise an index.

The core concept

An AI of truth, not an AI that guesses

A generative model produces what is plausible. ClaimTrace's AI does the opposite job: it opens the page, reads it the way a human would, and only serves words that exist in what it read.

It reads

JavaScript pages rendered in a real browser, linked PDFs, scanned images, event pop-ups — the real content, not an extract.

It anchors

Every value is tied to the sentence that proves it: URL, date, verbatim, confidence.

It abstains

A motivated not_found — absence is an answer, and it is documented.

Anti-hallucination for your AI agents

The honest null is what makes every other answer trustworthy

Invented ground data is the failure your users notice: a plausible menu, a plausible opening time, a plausible price. It costs more than an empty answer, because it is impossible to detect downstream.

✓Two states an agent can actually handle: a proven fact, or a documented absence.
✓A real partial fact is worth more than a complete invented one.
✓The absence carries the source that was read — it is auditable, not a shrug.

external adversarial audits

fabricated menus in the last 4 audits (91 entities in the last two, 31.07.2026)

3 299

automated tests green (full suite re-run 08.08.2026)

21/21

customer-journey assertions verified live on production (08.08.2026)

anchored golden cases re-checked before every deploy — a gap blocks and rolls back

Written into the engine

Eight guarantees, written into the engine

Not a policy page — four mechanical gates and eight guarantees a response either satisfies, or it is not served.

Dimension

The ClaimTrace guarantee

What you receive

A ClaimV1 object per fact.

Provenance

URL, date, verbatim, page kind, domain ownership.

When the data does not exist

A motivated not_found, never a guess.

Accepted sources

The entity's publications only — 17 aggregators rejected.

Homonyms

The address outweighs the name.

JS pages, PDFs, scans

Rendered and read, not skipped.

Freshness

Live revalidation: confirmed, contradicted, unverifiable.

For your AI agent

A proven answer or an explicit absence.

In thirty seconds

What you get

An AI that reads, never guesses

The page is opened and read. What was not read is not served.

Every fact carries its proof

Source URL, publication date, verbatim excerpt, confidence — in the response itself.

The honest null

not_found with its reason and the source that was read.

Real ground, not a summary

Today's menu, today's price — read where the entity published it.

Proven by adversarial audit

Eleven external audits designed to catch fabrication. The last four found no fabricated menu.

Swiss and sovereign

Built and operated by AXIONYX.io Sàrl, Neuchâtel. Public data only, traced provenance.

Request → response — nothing is dressed up

A real request, its real response

POST /v1/fetch

curl -X POST https://claimtrace.dev/v1/fetch \
  -H "X-API-Key: afk_…" \
  -H "Content-Type: application/json" \
  -d '{
    "thematique": "restaurants menú del día",
    "lieu": "Barcelona",
    "prompt": "the real menu of the day, with price if published"
  }'

FetchResultV1· abridged, prod 08.08.2026

{
  "schema_version": "1",
  "request_id": "req_463438843762408e826c",
  "entities": [ /* 6 entities — all honest not_found */ ],
  "verification": {
    "truth_rules_version": "2026-07-31.r17-menu-survit-au-silence-tous-chemins",
    "verified_live": true,
    "checked_at": "2026-08-08T19:23:58Z"
  },
  "usage": { "credits_charged": 5, "cache_hits": false },
  "attribution": ["© OpenStreetMap contributors (ODbL)"]
}

The honest null· Entre Dos, Barcelona · 08.08.2026

{
  "claims": [],
  "not_found": {
    "reason": "not_published",
    "detail": "Source officielle lue : la donnée
              demandée n'y est pas publiée.",
    "source_checked": "https://www.restaurante-entredos.com/"
  }
}

The detail sentence is served verbatim by the engine, in French — we display it as it comes. See the FAQ on response languages.

✓SSE streaming: entities arrive as they are verified.
✓24 h replay per request_id, isolated per key.
✓Measured 08.08.2026: ~3 s for an already-verified area, 64–163 s for a fresh collection with live verification.
✓Cache hits are served free of charge.

Credits for verified facts

Launch pricing — beta

1 credit = 1 entity freshly resolved and verified. A fact already verified recently (shared cache, at most 24 h for dated data) is served free of charge.

Discovery

0 CHF

50 credits on signup

50 credits
25 requests / day
Email support

Get a free key

Pack S

15 CHF

≈ 0.10 CHF / verified fact

150 credits
500 requests / day
Email support

Activate by email

Pack MRecommended

45 CHF

≈ 0.08 CHF / verified fact

550 credits
500 requests / day
Email support

Activate by email

Pack L

140 CHF

0.07 CHF / verified fact

2 000 credits
500 requests / day
Email support

Activate by email

Scale

99 CHF / month

your own AI key

Credits: on your own key
5 000 requests / day
Email support

Talk to us

Beta: packs are activated by email — write to us. Indicative launch pricing. Refusals (429, 402, 422) are never charged: the counter only counts admitted requests, and a 402 is returned before the request runs.

Before you integrate

Operational questions

How fast is a response?

Measured on 08.08.2026: an already-verified area answered in about 3 seconds; a fresh collection with live verification took 64 to 163 seconds depending on depth. The SSE stream sends entities as they come, so your agent starts working before the collection ends.

How are credits counted?

1 credit = 1 entity freshly resolved and verified. Cache hits are free. Refused requests (rate limit, empty balance, invalid input) are never charged.

What happens when my credits run out?

The API answers 402 before running the request — you are never charged for a refusal. Top up a pack (beta: by email) and the same key keeps working.

Are there daily limits?

Yes, per key and per UTC day: 25 requests/day on Discovery, 500/day on packs, 5 000/day on Scale. Beyond, the API answers 429 with a Retry-After.

What data is covered?

The audited vertical today is restaurants (menus, Switzerland first). Other topics run in generic beta mode: the response then carries an explicit warning that claims have no dedicated truth harness yet.

Which language are responses in?

Field names are stable English identifiers. Some human-readable details (e.g. the not_found.detail sentence) are currently served in French, verbatim from the engine.

Is this legal / compliant?

Public data only, traced provenance, Swiss nLPD and GDPR framing, OpenStreetMap attribution (ODbL). By design, no circumvention of any technical measure (login, captcha, paywall).

Can I replay a response?

Every response can be re-read for 24 h via GET /v1/requests/{id}, isolated per key.

All questions, honest answers →

Quickstart

In production in three requests

Get your key

One email, no card. 50 credits, displayed once.

export CT_KEY="afk_…"

Stream a request

Entities arrive one by one, each with its proof.

curl -N -H "X-API-Key: $CT_KEY" \
  "https://claimtrace.dev/v1/fetch/stream?thematique=restaurants%20menus&lieu=Neuch%C3%A2tel&prompt=le%20menu%20du%20jour"

Check your usage

Tier, daily quota, credit balance — the same values as your client area.

curl -H "X-API-Key: $CT_KEY" https://claimtrace.dev/v1/usage

✓Same key in the client area — no second account.
✓GET /v1/usage — balance and quota
✓GET /v1/stats — today's public service counters

Read the full quickstart → · API reference →

Why you can believe it

The proof is the audit

ClaimTrace exposes an engine already running in production behind arc-menu.ch. It has been through eleven external adversarial audits — audits written to catch it inventing. The last four found no fabricated menu; the last two alone covered 91 entities (31.07.2026).

Honest scope: the audited vertical is restaurant menus, Switzerland first. Other topics run in generic beta mode and the response says so explicitly.

✓Nine golden cases re-checked before every deploy — a gap blocks and rolls back.
✓3 299 automated tests green (full suite re-run on 08.08.2026).
✓21/21 customer-journey assertions verified live on production (08.08.2026).
✓17 aggregator sources rejected by rule — only the entity publishes about itself.
✓Versioned truth rules, carried in every response.
✓No circumvention of any technical measure — by design.

Ready to build

Get a free key

50 credits, no card, no sales call.

Your API key

Your key is displayed once, here. One key per email address.

Your key (copy it now — it will never be shown again):

Trouble? Write to kam@axionyx.io.