A mention records a source statement
Extraction produces mentions from versioned fragments. Each mention identifies its enrichment result and exact source fragment version.
A mention belongs to a dataset. Workspaces reaching that dataset can use the same eligible mentions. Source access still applies.
You cannot edit a mention. Correct the workspace’s interpretation through curation, or correct and reprocess the source.
Lists express the workspace’s current interpretation
Each workspace maintains canonical entities and relationships. Aliases connect mentions to these entries. Two similar names need not identify one entity. A list entry can be New, Accepted or Hidden. People can also add entries directly. Acceptance records a curation decision; it does not prove the underlying claim. The workspace pins a domain-schema version. That schema supplies layers, types, attributes, units and permitted relationship endpoints. Unlisted types appear under Other. Publishing another schema does not move the workspace’s pin. An Owner must select the new version deliberately.Evidence counts describe coverage
Several analyses of one item do not create independent observations. Several frames of one video still count as one observation. Counts do not establish independent corroboration.
Competing attribute values remain separate. A confidence score describes its stated method; it is not automatically a calibrated probability.