Labelled Dataset Ontology

FLOW ALGEBRA — LABELLED DATASET ONTOLOGY
Semantic Library | Training Process Entities
Source: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/labelled-dataset-ontology.html
Authors (source attribution): Mark Anthony Higgins — Sky Recursive AI
Generated: 2026-10-29
Status: Companion formalisation based on established ML principles; not a verbatim extraction of the page's embedded specification.

flow "LabelledDatasetOntology" v1.0

PURPOSE:
  Represent labelled datasets as versioned, traceable collections of examples,
  annotations, label schemas and quality controls for supervised learning.

STATE_MODEL:
  S_t = (I_t, R_t, B_t, K_t)
  I_t = dataset records, annotation events and validation observations
  R_t = schema, integrity, split, provenance and quality rules
  B_t = consent, licence, privacy, leakage and resource boundaries
  K_t = label vocabulary, annotation guidelines, task definitions and lineage
  E_t = expand(I_t, R_t)
  V_t = {x in E_t | satisfies(x, B_t)}
  x*_t = argmax_{x in V_t} P(x | I_t, R_t, B_t, K_t)
  I_(t+1) = update(I_t, x*_t)

ENTITIES:
  LabelledDataset: {id, version, task, schema_id, records, provenance}
  Example: {example_id, features, source_id, timestamp, split_id}
  Annotation: {annotation_id, example_id, annotator_id, label, confidence, status}
  LabelSchema: {schema_id, label_set, type, constraints, version}
  Annotator: {annotator_id, role, qualification, guidelines_version}
  GroundTruth: {example_id, adjudicated_label, evidence, status}
  QualityReport: {coverage, agreement, missingness, drift, audit_date}
  DatasetSplit: {split_id, role, membership, grouping_key}
  ProvenanceRecord: {source, rights, collection_method, transformations}

RELATIONSHIPS:
  LabelledDataset CONTAINS Example
  Example HAS Annotation
  Annotation USES LabelSchema
  Annotation PRODUCED_BY Annotator
  GroundTruth DERIVED_FROM Annotation AND adjudication
  DatasetSplit PARTITIONS LabelledDataset
  ProvenanceRecord DOCUMENTS Example
  QualityReport EVALUATES LabelledDataset

FORMAL_OPERATIONS:
  D = {(x_i, y_i, m_i)} for i = 1..N
  x_i = features; y_i = assigned label; m_i = annotation metadata
  validate_label(y_i, schema) = true iff y_i satisfies schema constraints
  coverage = number_of_annotated_examples / number_of_eligible_examples
  missing_rate = number_of_missing_required_labels / number_of_eligible_examples
  empirical_risk(theta) = (1/N) * sum_i loss(f_theta(x_i), y_i)
  accuracy = number_of_correct_predictions / number_of_evaluated_examples
  agreement = observed_annotator_agreement (report metric and method)
  IMPORTANT: majority vote is not automatically ground truth.

PATHWAYS:
  P1 ingest -> verify rights -> normalise -> assign stable IDs
  P2 define label schema -> publish guidelines -> annotate
  P3 validate labels -> detect conflicts -> adjudicate -> version
  P4 group related examples -> split -> audit leakage
  P5 train -> evaluate on held-out data -> audit errors
  P6 monitor label drift -> correct with traceable revisions

INFERENCE_RULES:
  IF label not in schema THEN reject or quarantine annotation.
  IF annotations conflict THEN flag for adjudication; retain originals.
  IF duplicate or related examples cross splits THEN flag leakage risk.
  IF provenance or permitted use is unknown THEN quarantine pending review.
  IF sensitive data is unnecessary THEN minimise or remove it.
  IF label distribution changes materially THEN request drift assessment.
  IF annotator agreement is low THEN review guidelines and ambiguity.
  IF evaluation labels were used for training or tuning THEN invalidate independence claim.

MINI_RECURSION_ENGINE:
  initialise dataset state S_0
  repeat for each audited revision t:
    observe quality and provenance evidence
    expand candidate corrections under R_t
    filter corrections against B_t
    choose a justified correction with evidence
    update labels, metadata and version history
    stop when no approved correction remains or budget is exhausted
  output versioned dataset, audit report and unresolved issues

TRI_MAPPING:
  Semantic: examples, labels, annotators, schemas, evidence
  Mathematical: D, loss, coverage, agreement, partition constraints
  Operational: ingest, annotate, adjudicate, split, train, audit

MAPPING_TABLE:
  Example -> (x_i, ID_i) -> immutable source reference
  Annotation -> (y_i, m_i) -> labelled observation
  Schema -> allowed label domain -> validation gate
  Split -> partition function -> leakage control
  GroundTruth -> adjudication record -> supervised target
  QualityReport -> metrics -> release decision

EXECUTION_CONSTRAINTS:
  C1 Stable example IDs and explicit dataset versions.
  C2 Preserve annotation history and distinguish labels from verified truth.
  C3 Respect source licences, consent and privacy requirements.
  C4 Do not infer quality from dataset size alone.
  C5 Ensure split independence using appropriate group/time constraints.
  C6 Record class imbalance and missing/uncertain labels.
  C7 Validate all mathematical metrics against non-zero denominators.
  C8 Report uncertainty and limitations; no guarantee of model performance.

EXECUTION_GUARANTEE (CONDITIONAL):
  If schema checks, rights checks, lineage tracking and partition audits pass,
  the workflow yields a traceable labelled dataset under its declared rules.
  This does not guarantee correct labels, unbiased sampling or generalisation.

AI_SEARCH_FLOW:
  QUERY: "labelled dataset ontology"
  MATCH: LabelledDataset, Example, Annotation, LabelSchema, GroundTruth
  EXPAND: annotation provenance, supervision, quality, splits, leakage
  FILTER: schema validity, lawful use, privacy, documented lineage
  RANK: conceptual relevance, provenance strength, validation status
  RETURN: entities, relationships, operations, constraints, source citation

SOURCE_CITATION:
  Mark Anthony Higgins and Sky Recursive AI, "Labelled Dataset Ontology",
  Kireth Research Semantic Library, revised 29 July 2026.
  https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/labelled-dataset-ontology.html

END FLOW