SOQVER Prepare a source

What works and what does not

Neither the website nor the documentation may claim more than this table. "Live" means you can use it today. "Tested" means it passes automated tests but nobody outside our team has used it yet.

live 24 · tested 6 · prototype 1 · planned 0 · as of 2026-11-21

Format and validation

WhatStateChecked
SO4AG 0.1 specification (working draft)live2026-11-21
JSON schema and profiles for vehicle rental and online shopslive2026-11-21
Validator on the web and on the command line
Accepts a file, JSON-LD or the address of a page with embedded JSON-LD.
live2026-11-21
JSON-LD form (schema.org and the so4ag vocabulary), lossless both ways
Every claim keeps its digest after conversion.
tested2026-11-21
Conformance cases for other implementations (32 for readers, 24 for validators)tested2026-11-21
Second reader in Python with no shared code
Written by the same team. No third-party implementation exists yet.
tested2026-11-21
Version 0.2 (working draft): file defaults, usage terms, live sources, a view with valid facts only
Every source has a short 0.2 file next to the main 0.1 file. The conversion is lossless both ways and both readers read it. 0.1 stays the main file; no decision to switch has been made.
tested2026-11-21

Creating a source

WhatStateChecked
Proposal of facts from the company website with a verbatim quote (rentals, shops)
A proposal is not a confirmation. Reading the same website twice does not always give the same result.
live2026-11-21
Free website audit: a score in five areas and specific findings, the result after registering with an e-mail
Since 21 November 2026. It measures whether facts can be found, are unambiguous, current and machine-readable. It does not say whether AI will recommend the business. Facts are extracted only for shops and vehicle rental; other industries get the audit without them.
live2026-11-21
Confirmation by the company: a code sent to an address on its domain
Proves access to a mailbox on the domain, not full authority to act for the company.
live2026-11-21
Manual entry, CSV import, change history and revisionslive2026-11-21
Import of delivery prices from the shop’s product feed (Heureka, Google Merchant)
In the admin of online shops: the feed generated by the shop system becomes drafts for approval. Checked on three real public feeds: delivery and cash-on-delivery prices matched what we had read from the websites of those shops. No company has used it in the admin yet. There is no direct connection to a shop back office.
live2026-11-21

Publishing

WhatStateChecked
Free package for your own website: so4ag.json on your domain, a Facts and terms page, llms.txt, instructions
At most 25 facts, valid for 60 days at most, renewed by hand. The company confirms the facts by ticking a box; we do not verify its identity for the download. The placement check only verifies that the file is on the domain and is valid.
live2026-11-21
"Facts and terms" page hosted by us with embedded JSON-LDlive2026-11-21
so4ag.json, JSON-LD and Markdown of every sourcelive2026-11-21
The Markdown form separates facts valid now from announced ones and from those no longer valid
Since 21 November 2026. In the comparison test models erred most where an old and a new value stood side by side. so4ag.json and JSON-LD still carry every claim of the revision.
live2026-11-21
Export for your own website: finished page, block to paste, file with identifiers on your domain
Public forms can be downloaded without an account.
live2026-11-21
Check of a self-hosted copy against the latest revision, with the signature
On request and once a day automatically. When the company website stops matching, a finding is opened.
live2026-11-21
Signed revision: an Ed25519 signature and a public key for every confirmed source
The signature shows the file has not changed since publication. A key on the company domain shows the key is the company’s; a key only on our site shows only publication through the company account. It does not show the facts are true.
live2026-11-21

Reading

WhatStateChecked
Public API and MCP, read onlylive2026-11-21
Reading rules: scope, validity, "not stated" is not zero, a "from" price stays a minimumtested2026-11-21
Readers told apart: our own reads, AI crawlers, other programs, browsers
The class comes from the program header and can be forged. We do not yet have a reader outside our team who demonstrably used the facts in an answer.
live2026-11-21

Checks and trust

WhatStateChecked
A quote with every proposal and plausibility ruleslive2026-11-21
Company identity from the public register (ARES)live2026-11-21
Reviewer attestation bound to the exact wording of a claim
The format and validator support it. No attestation has been issued yet.
tested2026-11-21
Regular check that the quoted evidence is still on the company page
Every 7 days, every 14 for unconfirmed proposals. When the quote disappears, a finding is opened for review. It checks the wording, not the meaning, and only evidence that has a quote.
live2026-11-21
Register of findings: reports from readers, decisions, disputed facts, appeals
The form is under every source. A person decides. An accepted finding marks the fact as disputed in every form. No report from outside has arrived yet.
live2026-11-21
Links to review profiles in the source
The company lists its review profiles and the source carries them. We do not copy ratings or texts. There is no data link with BRANDprotecto or ReviShield, only a link from the admin.
live2026-11-21

Measurement

WhatStateChecked
Counter of views and machine reads without cookies or visitor addresseslive2026-11-21
Measurement in four steps: published, connected, used in an answer, contributed to an outcome
In the admin of every source, each step with what it shows and what it does not. We cannot see an order or a booking; the last step is a click through to the company website.
live2026-11-21
Comparison test: the same facts as a page, JSON-LD and SO4AG, with and without a reader
Two runs, three models: online shops (242 tasks) and vehicle rentals with larger sources (397 tasks). With a reader every task came out right in both notations. Without a reader small sources showed no difference (98 to 100 %), larger ones 85 to 100 % depending on the model, and the plain page was read best. Most errors come from old and new values side by side. No benefit of the custom notation over JSON-LD with a reader was shown. The larger sources are test fixtures and a third industry is missing.
prototype2026-11-21

What nobody has shown yet

Machine-readable form of this table: /spec/capabilities.json