Assessment
- CURIE:
gmeow:Assessment - IRI: https://blackcatinformatics.ca/gmeow/Assessment
- Category: class
- Defined by:
gmeow:slices/norms - Box roles: TBox role (What is this?)
An observation that scores a target against a criterion (or whole rubric) — vantage = the judge, human or model: an LLM judge is just a vantage, and two models disagreeing are two coexisting cells with no winner (Principle 9). The judge model's identity and prompt digest travel as observation provenance. gmeow:assessmentCriterion / gmeow:assessmentRubric play the observationMethod ROLE but are deliberately not subproperties of it (observationMethod is functional with a QualityValue range; Criterion and Rubric are Entities — the claimModality axiom pattern). Scoring-density expectations by target granularity reuse kernel gmeow:hasGranularity on the target.
Structure
Subclass of: gmeow:Observation
Practical Pattern
Use gmeow:Assessment as a specialized kind of gmeow:Observation. Add statement metadata or a standpoint when the assertion needs provenance, confidence, or vantage.
Common Companion Terms
Projects To
| Profile | External Targets |
|---|---|
schema-org |
rdf, schema |
Linkages
Generated from the canonical mapping DSL. SSSOM files are the generated public interchange form for term equivalences.
Projection Coverage
| Source | Kind | Profile | Predicate/Relation | Target | Evidence |
|---|---|---|---|---|---|
gmeow:Assessment |
projection | schema-org |
projects to / <= | rdf:type, schema:ClaimReview, schema:Rating, schema:author, schema:itemReviewed, schema:ratingValue, schema:reviewRating | gmeow:mapSchemaClaimReview; confidence 0.85; lossy: the criterion/rubric structure and scale metadata collapse to a bare ratingValue; each review stays authored by its vantage — N competing assessments emit N coexisting ClaimReviews, never one aggregated rating (P9) |
Usage Advice
Use when
- Use for a score of a target against a criterion or rubric — an observation whose vantage is the judge, human or model; disagreeing judges are coexisting cells with no winner.
Avoid when
- Avoid treating a model's score as ground truth (an LLM judge is just a vantage — Principle 9) and avoid omitting a zero score; a zero is a score, not an absence (false positives are worse than zeros).
How to use
- Mint one
Assessmentper criterion (a 21-axis pass is 21 Assessments), naming the target withgmeow:assessmentTarget, the criterion/rubric withgmeow:assessmentCriterion/gmeow:assessmentRubric, the score withgmeow:assessmentScoreValue, and the judge as observation vantage.
Examples
- ex:as1 a
gmeow:Assessment;gmeow:assessmentTargetex:chapter3;gmeow:assessmentCriterionex:agency;gmeow:assessmentScoreValue0.9.