From a passage to a proposed concept
A technical view of source-grounded extraction, concept roles, review, work identity and the boundary between a description and an executable control.
Implemented extraction and integrity checks · semantic acceptance remains open
Extract a proposal, retain the source.
The local discovery implementation selects passages from training records and asks a pinned instruction model for structured concept proposals. Each proposal includes a name, definition, role and exact supporting quotation. Operator proposals additionally require input state, result state and applicability.
Validation checks the schema and whether the quotation occurs exactly in the source passage. These are attribution and integrity checks. They do not establish that the source entails the proposed definition; that is a separate semantic review question. Held-out artwork descriptions are not used to grow the extraction vocabulary.
Keep different kinds of knowledge distinct.
The research distinguishes schools, properties and operators. A school is a historical grouping; a property describes a characteristic; an operator describes a transformation with an explicit scope. A movement name cannot silently become a brush setting.
Claims retain their target work, source, modality and relation, including qualification or disagreement. A sentence about an artist or a period does not automatically become evidence for every painting. Related concepts need not be synonyms, and lexical normalization creates review proposals rather than merging the registry by itself.
Connect a claim to an experiment explicitly.
The control-evidence implementation checks concept, ontology, work and reference identities before describing a mapping. Only selected work-scoped concepts have explicit connections to the current full-work painting path. Other concepts may remain descriptive or belong to different experiments.
The broader Picasso platform also contains artwork enrichment and entity-resolution infrastructure. That infrastructure and the experimental painting ontology have different records and acceptance boundaries. A research claim is not automatically promoted into a production graph or made available as a public painting control.
What the records carry
| Record | What it carries | How to interpret it |
|---|---|---|
| Extraction candidate | Definition, role, exact quotation and transformation fields where applicable | A well-formed proposal still needs semantic review. |
| Work-scoped claim | Target identity, modality, source span and support or qualification | One work does not establish an artist-wide invariant. |
| Control binding | Exact ontology, concept, work and image identities plus a rendering mechanism | The mechanism does not prove aesthetic or historical validity. |
Methods summary from the inspected research snapshot. Findings described here belong to their stated experiments; this note is not a fresh replication or a peer-reviewed publication.