A correct number can still produce a false claim when its denominator changes.
Consider:
Electric cars accounted for 18% of the global car fleet in 2023.
The IEA Global EV Outlook 2024 attributes approximately 18% to new-car sales, not the total fleet. That distinction matters when an agent uses the figure in a market report.
This article shows how SAVER Verifier checks a selected claim and returns evidence linked to source URLs.
Disclosure: We operate SAVER Verifier. The production test described here was operator-funded, not a customer purchase. This article was prepared with AI assistance and reviewed against the recorded test.
Inspect the API without paying
A valid unsigned request returns an x402 payment challenge. This command does not authorize payment:
curl -i https://api.saververify.com/v2/verify \
-H "Content-Type: application/json" \
--data '{"claim":"According to IEA Global EV Outlook 2024, electric cars accounted for 18% of the global car fleet in 2023.","context":"Check the metric and denominator against https://www.iea.org/reports/global-ev-outlook-2024/trends-in-electric-cars"}'
Expect HTTP 402, with payment requirements encoded as base64 JSON in the PAYMENT-REQUIRED header. The response body may be {}.
The current price is 0.20 USDC per completed verification on Base Mainnet. An authorized x402 wallet is required to purchase a check; inspecting the challenge requires no wallet.
What the paid test returned
Our 26 October production test returned contradicted.
The explanation identified the difference between annual sales and fleet stock. The assessment linked its evidence to these IEA pages:
The response uses:
-
status: supported, contradicted, mixed or insufficient_evidence -
explanation: the assessment and relevant qualifications -
evidence: findings with a relation andsourceIds -
sources: the corresponding source IDs and URLs
Resolve each evidence item's sourceIds against sources.id. A working URL alone does not establish that a source supports the claim.
Purchase one check safely
The public repository includes a bounded reference client:
- Save your claim and optional context in
request.json. - Inspect the request and payment terms.
- Authorize one purchase through your runtime's existing wallet and spending policy.
- Validate both the returned assessment and settlement receipt.
The client instructions are here:
Never paste wallet secrets into a model conversation. A timeout or lost response is not permission to sign another payment. Opt-in result recovery retains settled results for 24 hours; the reference client saves the recovery information needed to reconcile an attempt.
Use it as a report checkpoint
An agent can select one material claim before delivering a report, request a check, and retain the assessment alongside that passage.
Checking one passage does not verify the whole report. A supported assessment also does not authorize publication or bypass the agent's existing policies.
The repository includes a free stored-result demo:
node examples/research-report-review.mjs --demo
Run it from a checkout of the public repository, using Node.js 18 or later. It replays an existing operator-funded record locally; it makes no network request, model call or new payment.
What this does—and does not—demonstrate
SAVER provides a ready-made API with structured assessments, evidence links, payment handling and result recovery.
It is not a guarantee of factual correctness. Confidence values are model-assessed, not calibrated probabilities.
In our small four-case comparison, direct GPT-5.4 with web search reached the same correct main conclusions as SAVER and had lower estimated model cost. These tests therefore do not establish higher accuracy or lower cost than building your own checker. The integration is relevant when you want the packaged workflow rather than maintaining it yourself.
You can inspect the stored paid examples for free, or submit a claim from your own report using the documented payment controls.
Top comments (0)