Synthetic Dataset Ontology

FLOW ALGEBRA — SYNTHETIC DATASET ONTOLOGY
Semantic Library | AI Core > Entity Graph > AI Entities > Training Process Entities
Source: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/synthetic-dataset-ontology.html
Authors: Mark Anthony Higgins — Sky Recursive AI
Document type: Companion formalisation; not a verbatim transcription of the embedded page specification.
Language: en-GB | Encoding: UTF-8 | Generated: 2026-10-29

flow "SyntheticDatasetOntology" v1.0

PURPOSE
  Represent the creation, governance, verification and use of synthetic datasets
  as typed entities and explicit state transitions. Synthetic means generated
  or transformed rather than directly collected as the target observation.

STATE_MODEL
  S_t = (I_t, R_t, B_t, K_t)
  I_t = dataset records, generator configuration, seeds, reference distributions
  R_t = generation, filtering, labelling, splitting and validation relations
  B_t = privacy, safety, provenance, quality and resource constraints
  K_t = accepted facts, audit trail, metrics, limitations and known risks
  E_t = expand(I_t, R_t)
  V_t = {x in E_t | satisfies(x, B_t)}
  a_t = argmax_{x in V_t} P(x | I_t, R_t, B_t, K_t)
  I_(t+1) = update(I_t, a_t, observations_t)
  Note: argmax is a selection formalism, not a calibrated probability guarantee.

ENTITIES
  SyntheticDataset: versioned collection of generated records
  SyntheticRecord: one generated example with an immutable record identifier
  Generator: simulator, procedural model, learned model or augmentation process
  GeneratorVersion: immutable identifier of generator code and weights
  Seed: reproducibility input for stochastic generation
  GenerationPrompt: optional instruction or conditioning input
  SourceDataset: optional real or synthetic conditioning data
  ReferenceDistribution: target properties or distribution for evaluation
  Label: target or annotation, generated or independently verified
  ProvenanceRecord: generator, source, licence, timestamp and transformation chain
  QualityMetric: measured fidelity, diversity, coverage or validity
  PrivacyAssessment: documented disclosure and memorisation risks
  SplitAssignment: training, validation or test role
  HumanReviewer: authorised assessor of generated records
  DatasetRelease: versioned, approved export and associated documentation

RELATIONSHIPS
  SyntheticDataset contains SyntheticRecord
  Generator produces SyntheticRecord
  Generator uses Seed and optional GenerationPrompt
  SyntheticRecord derives_from optional SourceDataset
  SyntheticRecord has Label and ProvenanceRecord
  SyntheticDataset evaluated_against ReferenceDistribution
  QualityMetric assesses SyntheticDataset or SyntheticRecord
  PrivacyAssessment constrains DatasetRelease
  SplitAssignment assigns SyntheticRecord to an evaluation role
  HumanReviewer verifies selected SyntheticRecord

FORMAL_OPERATIONS
  D_syn = {G(z_i, c_i; theta_G) | i = 1,...,N}
  where z_i is random input, c_i is optional conditioning information,
  and theta_G denotes generator configuration or parameters.
  p_syn(x) = distribution induced by G and its input distribution.
  FidelityGap = d(p_syn, p_target) for a documented distance d.
  Coverage(A) = count({x in D_syn : x in A}) / max(1, |D_syn|).
  Acceptance(x) = Valid(x) AND ProvenanceComplete(x)
                  AND PolicyCompliant(x) AND QualityPass(x).
  D_accepted = {x in D_syn | Acceptance(x)}.
  DuplicateRate = duplicate_records / max(1, total_records).
  MemorisationRisk = estimated likelihood of disclosing training examples;
                     must be measured, not assumed absent.
  The generator's labels are not automatically independent ground truth.

PATHWAYS
  P1: Specify task, population, coverage goals and intended use.
  P2: Register generator, licences, inputs, seeds and configuration.
  P3: Generate candidate records with per-record provenance.
  P4: Validate schemas, constraints, labels and domain plausibility.
  P5: Measure distributional fidelity, diversity, duplication and bias.
  P6: Assess privacy, leakage, safety and source licensing.
  P7: Partition by provenance groups before model selection/evaluation.
  P8: Approve versioned release with dataset card and limitations.
  P9: Monitor downstream performance and revise or revoke if needed.

INFERENCE_RULES
  R01 IF provenance is missing THEN quarantine record.
  R02 IF validation fails THEN reject or regenerate record.
  R03 IF synthetic data resembles protected source examples too closely
      THEN flag for privacy and licensing review.
  R04 IF train and test records share source-derived information
      THEN flag potential evaluation leakage.
  R05 IF fidelity is high but coverage is low THEN do not infer completeness.
  R06 IF labels come solely from the generator THEN mark as unverified.
  R07 IF distribution shift exceeds tolerance THEN require re-evaluation.
  R08 IF safety or privacy approval fails THEN block release.
  R09 IF generation is stochastic THEN record seeds and environment details.
  R10 IF performance improves only on synthetic evaluation data
      THEN require independent real-world or held-out assessment.

MINI_RECURSION_ENGINE
  initialise specification, generator, constraints and target metrics
  repeat until acceptance target reached or resource limit exceeded:
    candidates <- generate(generator, seed, conditions)
    valid <- filter(candidates, schema, provenance, policy)
    metrics <- evaluate(valid, fidelity, coverage, privacy, leakage)
    if metrics satisfy release thresholds: stop and request approval
    else revise generator or sampling plan; log all changes
  terminate on acceptance, maximum iterations or safety failure
  No convergence or quality improvement is guaranteed.

TRI_MAPPING
  Entity: SyntheticRecord
  Relation: produced_by(Generator), has_provenance(ProvenanceRecord)
  Process: generate -> validate -> audit -> approve -> publish

MAPPING_TABLE
  Generator           -> production mechanism -> reproducible generation
  SyntheticRecord     -> data instance        -> validity and traceability
  ProvenanceRecord    -> evidence link        -> audit and lineage
  QualityMetric       -> evaluation evidence  -> fidelity and coverage
  PrivacyAssessment   -> release constraint   -> disclosure risk control
  DatasetRelease      -> output artefact      -> versioned approved dataset

EXECUTION_CONSTRAINTS
  All records must have unique IDs and explicit schema versions.
  Every release must identify generation methods and permitted uses.
  Do not assert that synthetic data is automatically private or unbiased.
  Never mix generated labels with verified labels without provenance tags.
  Use disjoint, lineage-aware splits to reduce evaluation contamination.
  Maintain documented uncertainty and human-review requirements.
  Reject invalid, unsafe or unauthorised outputs.

EXECUTION_GUARANTEE
  If every required predicate is evaluated correctly, only records passing
  the declared acceptance predicates enter D_accepted. This does not prove
  factual correctness, privacy, fairness or real-world generalisation.

AI_SEARCH_FLOW
  QUERY: synthetic dataset ontology, generated training data, data provenance
  IDENTIFY: synthetic data type, generation process, intended application
  EXPAND: generator -> records -> labels -> validation -> release
  FILTER: schema, provenance, quality, privacy, licensing and leakage
  RANK: relevance to query, verified evidence and applicability
  RETURN: definitions, formal relations, caveats and source citation
  SOURCE: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/synthetic-dataset-ontology.html
  CITATION: Mark Anthony Higgins and Sky Recursive AI, Kireth Research

END FLOW