Model and system cards
A card is a structured description of an AI model or an AI system, written for the people who review, approve, audit or use it.
- A model card follows Model Cards for Model Reporting (Mitchell et al., 2019): model details, intended use, factors, metrics, evaluation data, training data, quantitative analyses, ethical considerations, and caveats and recommendations, plus a section on compute and environmental impact. Each section has its own data points (the facts a reviewer looks for by name, such as who developed the model, its size, its inputs and outputs, where its training data came from and whether it contains personal data) and a free-text narrative.
- A system card covers system details, intended use, out-of-scope uses, affected people, human oversight, data, performance and monitoring, and risks and limitations. It also includes the latest published card of each model the system uses.
Drafting
A system's card can be edited by anyone who can edit the system. A model's card can be edited by anyone except auditors.
A section you leave empty is filled from the registry: the system's intended use, its oversight design, a model's training-data summary, and so on. Write a section yourself to replace that text. The From the registry panel shows what the card draws on; facts the registry owns (a model's provider, license, modality, hosting and limitations) are edited on the model's own page. Developed by shows the registry's provider until you type a value, which then replaces it on the card (for a model one vendor develops and another hosts).
The section list beside the draft shows how much of each section is recorded and marks the sections whose content differs from the latest published version (an edit to the registry record counts, because publishing would freeze it). A value that still says [To confirm is counted as not recorded, and the page says how many remain. A data point you leave empty is left off the published card, except Developed by, which publishes the registry's provider. The General-purpose AI model as defined in the EU AI Act data point is your organization's own determination, and is published as that. Recording it does not decide what the EU AI Act asks of your organization: the Act's obligations for general-purpose AI models (Chapter V) fall on the model's provider, and whether your organization is a provider depends on the Act's own definition (Article 3(3)), not on this data point.
Research with Watney
On a model card, Research with Watney looks up the model's public documentation (when web search is on) and works through the card with you in a conversation. Research with Watney can propose any data point except the version the card describes and the EU AI Act determination, each with where Watney found it; nothing it proposes reaches the card until you apply it and save. Only a value from a page Watney's searches returned is ticked for you, and only for an empty field. A value whose source is marked Source not verified cites a page Watney's searches did not return: check it before you apply it; once applied, the field keeps that label until you edit the value yourself (the label is not saved with the card). Watney also asks about what no source states; answer in place to keep going. Your recent conversations about the model are listed above the conversation, so a reload does not lose one: Reopen shows each request and Watney's answer as they were recorded, and your next request continues it, up to the conversation's limit of requests. Only you see your own conversations. On the Add a model page, Look up a public model fills the form from the developer's documentation, and nothing is added until you add it.
Backsplice deletes a research conversation once 12 months have passed since its last request, every request and answer in it together, and a lookup once 12 months have passed since it was made; a nightly job does it. A value you applied to the card is part of the card and stays. Your organization's audit log keeps that each research request and lookup was made, without what was asked. Copies remain in Backsplice's backups until those expire. Watney's card page names the model provider and the models in use; its system card, which will describe how the provider handles what is sent, is not published yet.
Research with Watney, Draft with Watney and Look up a public model appear only when your organization has turned Watney on and your plan includes it; without them every field can still be written by hand.
Publishing
Publishing freezes the card as a numbered version and records who published it and when. A published version never changes. Edit the draft and publish again to make a new version.
Open any published version from the Published versions list to read it in the app, with its integrity check, without downloading it.
Each version carries a tamper-evidence check, shown as Intact. It shows whether the version has been altered without the application's signing key since it was published. Publishing a card records who published it; it is not an approval of the system, which only a review decides.
Save and publish saves what is on screen and publishes it in one step, so an unsaved edit is never lost.
Draft with Watney
If you can edit the card, Draft with Watney drafts every section from the record: for a system card, the system's registry entry, its models, its open risks and a count of its incidents by type; for a model card, the model, its versions and its evaluation results. You can paste the relevant parts of the vendor's documentation for it to use. On a model card it also fills the descriptive data points you have left empty, never one you typed, and never a date, a link, a choice or a determination. The draft fills the fields on screen and is not saved: read and correct every section, then save or publish it yourself. Where the record is silent, the draft says [To confirm: …] instead of guessing, and Watney lists what the record does not say. If you have already typed in the fields, you are asked before they are replaced. A card that still says [To confirm anywhere cannot be published: state the fact or remove the sentence first, because a published version never changes.
Published versions can be downloaded as Markdown, JSON or PDF once your organization is on a paid plan (downloads are not on the Free plan). The latest ones are included in the system's approval dossier.