Self-assessment

Select the statement that best describes the curation team today, one per chapter. The result shows a level for each chapter and the chapters to read first. Answers are stored in this browser only.

1. Scope How is model assistance scoped?

Not yetStaff use assistants informally, with no rule on what may be sent or published.FoundationalA written rule names the tasks models may assist, forbids sending confidential records to external services, and requires curator approval before publication.AI-readyA task typology with decision rights is in place, and the records to start with are chosen from use and completeness.AI-nativeThe scope is reviewed from acceptance rates and incidents each period.

2. Migrate How are legacy records moved into the schema?

Not yetBy retyping, or they stay where they are.FoundationalColumns of one source are mapped by hand and migrated with a script; the report is reviewed.AI-readyA model proposes the mapping with confidence and notes, a curator confirms each row, every source is migrated, validated, and routed into review.AI-nativeMappings are versioned per source, re-run on change, and unplaced columns feed schema change requests.

3. Draft How are fields drafted?

Not yetBy hand, or with ad hoc prompts.FoundationalA fixed prompt with the record and the schema field definitions produces drafts marked as drafts.AI-readyDrafts are structured output checked against the schema and the completeness profile, with acceptance rates measured per field.AI-nativeDrafting runs over the catalog with templates per field, and acceptance rates gate prompt changes.

4. Assess How is metadata quality assessed?

Not yetNot systematically.FoundationalCurators score a sample on a rubric.AI-readyA model scores records on fixed dimensions with structured output, calibrated against curator scores on a sample.AI-nativeScores are aggregated into catalog reports each period and drive the work list; calibration is re-checked when the model changes.

5. Review How are suggestions reviewed?

Not yetSuggestions are applied or ignored without a record.FoundationalA curator reads suggestions and applies the accepted ones by hand.AI-readyA review pipeline produces issues with category, severity, and a proposed fix; curators decide in a board with a diff view; decisions are recorded with reasons.AI-nativeDecisions flow back to the catalog through the editor, and to the evaluation set.

6. Standardize How are keywords and topics standardized?

Not yetFree text; each curator writes their own.FoundationalA list of preferred terms exists and curators consult it.AI-readyA SKOS vocabulary with alternates and URIs is maintained; free text is mapped to it, with unmapped terms reviewed.AI-nativeMapping runs over the catalog with model proposals for unmapped terms, and labels exist per language under native review.

7. Translate How are records translated?

Not yetThey are not, or the browser translates the page.FoundationalNames and keywords of the most used records are translated with the vocabulary's labels and reviewed by a native speaker.AI-readyDefinitions and units are translated by a model with the glossary, checked automatically, reviewed, and published per language with status and reviewer.AI-nativeAcceptance per language and field is tracked, changes are gated on reviewed translations, and new records are translated at publication.

8. Improve How do decisions improve the system?

Not yetThey do not; prompts change when someone has an idea.FoundationalAcceptance rates are reviewed occasionally.AI-readyAcceptance rates and an error taxonomy are tracked per field and model; prompt and model changes are gated by an evaluation set built from decisions.AI-nativeRoutine tasks run on small or local models chosen by the evaluation; cost per record is known.

0 of 8 answered

Interpretation of the result

• Chapter 1 first: the rule on confidential records and on curator approval takes a day and prevents the costliest mistakes.
• Chapter 2 where the records are in spreadsheets: nothing else can run on them until they are in the schema.
• Chapter 5 before scaling chapters 3 and 7: a drafting or translation run without a review board and recorded decisions produces text nobody checked.
• Chapter 8 is where the investment pays: without recorded decisions, every prompt change is a guess.