Curation and AI

MatSci-SAM extends the vocabulary workflow evaluated in MatSci-YAMZ (Greenberg et al., 2025). Its human-in-the-group AI workflow combines contributor review with model drafting. Human, model, and simulated activity receive distinct attribution.

Contributor and administrator roles

Contributors create terms and definitions, add examples, and manage topics on their own definitions. New terms use the selected community vocabulary or the default vocabulary under Everything. A topic creator can link the topic to an equivalent term when the scheme permits it.

Administrators assign facets, edit tag definitions and scope notes, and merge tags. Collection changes follow the stored membership policy. Administrator collections accept administrator changes. Contributor collections accept changes from signed-in contributors when that creation mode is enabled. It is disabled by default.

Application assertions record author and time. Retractions preserve those facts and add the retracting author and time. Migrated tagging records may lack an author because the earlier tables did not store one.

Tags and semantic change

A tag retains its identifier and preferred label. Administrators can edit its definition, scope note, and alternative labels. A semantic replacement uses a merge. The original is retired, its identifier redirects, and active statements are retracted and asserted for the replacement. Both records remain.

An equivalent-term link follows the definitions of the linked term. Administrators can retract the link if the meaning no longer fits the tag. The Tag drift report lists linked tags whose term definitions changed by at least 25 percent after classification began.

Contribution actions and language-model drafting

The five contribution actions create a new term, suggest a revision, propose a replacement, comment, or add an example. Model drafting is confined to New term and Suggest a revision. Both generate definition text only.

The application stores the draft, prompt, and model tag before the contributor reviews it. Publication attributes the definition to the contributor and model. A suggested revision creates a separate candidate linked to the exact source revision and critique. Discarded drafts remain outside the vocabulary.

Examples have independent contribution records. A featured example controls compact display while the full active set remains in the metadata. Example selection leaves definition revisions and votes unchanged.

Model profiles at /models/{model} identify a runtime tag, inferred publisher and family, parameter size when available, directly authored definitions, and recorded prompts. Model-account display names begin with MatBot. The tag records the requested model configuration.

Studies and support

A study records ordered participation over a collection. A Position accepts an existing definition, publishes an AI-assisted alternative, or proposes a new definition. A participant may also skip before contributing.

Accept records or retains an upvote. A published proposal completes Position without a vote. A suggested revision names its source revision. An independent proposal has no derivation or replacement target.

ID4 round two permits comments during Position and omits the repeated Review round. Other studies may have Review steps for additional votes and comments. See Studies.

Contributions made in the activity identify their step. Completions and closing answers are separate records. Vocabulary support can include votes from outside the study, and the overview does not publish it as consensus.

Participation requires an active membership episode and an open study. The RDF describes the study window and collection, while rosters and invitations remain private application data.

Simulated participants use separately labeled accounts. Generated definitions, comments, and text answers record their model tag and prompt. Votes, comments, and answers identify the simulated actor kind. The provenance model specifies their export.