Frequently asked questions

There is no public sign-up. If you already collaborate with us and need an account, write to kontakt@vetcaseaudit.de.

No, and it is not meant as a stage before it either. VetCaseAudit screens reporting completeness: whether a manuscript reports what a reporting checklist asks for. It does not judge whether the study was well done, the statistics correctly calculated or the conclusion sound. A case report can satisfy all 51 items and still be scientifically weak, and the other way round.

The file, the text extracted from it, the analyses and the supporting quotes sit on our server. What goes to the language-model provider is only chunks of the manuscript text and the checklist items, not the file, not the account data. You can delete at any time. The associated analyses and supporting quotes go with it.

The detailed data flow is under Security. The binding document is the privacy statement.

The provider is GWDG SAIA: anyone using the application stores their own access key for it, and the usage relationship is therefore between the organisation that uses it and that provider. Which of that provider's models are available is held in the application and maintained by us. Which one a particular analysis uses, you decide when you start it.

Every report states which provider and which model actually ran. Analyses that run on your own key are not billed by us.

The analysis fails visibly. There is no silent failover to another provider or another model, and there is no partial report from the chunks that did get through. You get an error message and decide about re-running it yourself.

The reason is set out in full under Method: a report in which failed chunks were counted as “missing” would look like a manuscript with gaps when the gap was in the procedure.

The report is a working basis, not a clearance. It makes no scientific judgement, its verdicts are AI-generated and can be wrong, and every finding carries its supporting quote so that it can be checked before anyone acts on it. Responsibility for the manuscript, and for what is made of the report, stays with the authors.

Yes, entirely, and it is labelled as such. Every verdict in the report comes from an AI system. In the application the notice sits directly at the score, under every generated text suggestion, and in the header of the PDF export, so that it travels along even when only a single page is passed on.

That is not a courtesy but the transparency obligation under Art. 50 of the AI Act, applicable since 2 August 2026. A model identifier in a footer is an evidential detail for someone who already knows what it means, not a label.

No. There is no connection to Shibboleth, SAML or identity management, and there is no batch mode across many manuscripts at once. One document, one checklist, one report. Access runs through invitation and password.

Not in version 1. Checklists are versioned data rather than hard-wired logic. Adding another means writing a file and publishing a version, but the authoring and release happen at our end. There is no interface in which users create their own checklists, and none is planned for version 1.

All members of the same organisation in the application. It is a shared workspace, not a set of separate mailboxes: an uploaded document is visible to the other members, and every document shows who uploaded it. Anyone who does not want that does not upload the document into this organisation.

Documents and the analyses and supporting quotes attached to them can be deleted by you at any time, individually as well. There is currently no automatic deletion period. Deleted data can persist in the backups for up to 14 days, because that is how long they are kept.

The complete retention periods are in the privacy statement, section 6.

It is auditable, which is a different thing. Every report carries: provider, model identifier, checklist key and version, generation parameters, number of chunks, token counts and timestamps. Those details travel with the PDF and the JSON export, and the checklist version is immutable. A later update does not change an existing report.

What does not follow from that: that a second run produces the same report. A language model's verdicts are not deterministic. What is citable is therefore what actually ran, not a result that could be reproduced at will.

German-language manuscripts have not been tested. All test runs so far used English-language texts.

Text extraction and server-side quote verification operate on characters and are not tied to a language. The checklist items and the instructions to the model are in English, however. Which language a model then writes its rationales in has not been checked. Suitability for German-language manuscripts therefore cannot be confirmed at present.