Demonstration Dataset Ontology

FLOW ALGEBRA — DEMONSTRATION DATASET ONTOLOGY
Version: 1.0 | Date: 2026-10-29 | Encoding: UTF-8
Authors: Mark Anthony Higgins — Sky Recursive AI
Source: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/demonstration-dataset-ontology.html
Status: Companion formalisation; embedded source specification not available for verification.
Purpose: Define the entities, relations, transformations and validation constraints for demonstration datasets used in imitation learning and supervised AI training.

01. STATE MODEL
S_t = (I_t, R_t, B_t, K_t)
I_t: observed demonstration examples and metadata.
R_t: schema, ordering, provenance and task relations.
B_t: validity, safety, licensing, privacy and split boundaries.
K_t: known tasks, policies, annotation conventions and evaluation criteria.
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, feedback_t)
Note: probabilistic ranking is optional; it never overrides hard validity constraints.

02. ENTITY TYPES
DemonstrationDataset: versioned collection of demonstrations.
Demonstration: recorded example of desired behaviour on a task.
Task: objective and admissible actions or outputs.
Context: initial conditions, instruction, observations and constraints.
Trajectory: ordered sequence of steps (optional for single-turn examples).
Step: observation, action, optional feedback and timestamp/order.
Expert: human, validated system, or other identified demonstration source.
TargetOutput: desired answer, action, or completed artefact.
QualityLabel: assessment with documented rubric and annotator.
ProvenanceRecord: source, collection method, permissions and lineage.
SplitAssignment: training, validation or test membership.
DatasetVersion: immutable snapshot identifier and manifest.

03. RELATIONSHIPS
Dataset contains Demonstration.
Demonstration addresses Task.
Demonstration beginsWith Context.
Demonstration hasTarget TargetOutput.
Demonstration mayContain Trajectory.
Trajectory orders Step by step_index.
Step observes Observation and emits Action.
Demonstration producedBy Expert.
Demonstration hasProvenance ProvenanceRecord.
Demonstration assessedBy QualityLabel.
Demonstration assignedTo SplitAssignment.
Dataset identifiedBy DatasetVersion.

04. TYPED REPRESENTATION
D = {d_1, ..., d_N}; N >= 0.
d_i = (id_i, task_i, context_i, target_i, trajectory_i?, metadata_i).
trajectory_i = [(o_1,a_1),...,(o_T,a_T)] where T >= 1.
Unique(d_i.id) within a dataset version.
Each action must belong to the task's permitted action space.
Single-turn demonstrations may have no trajectory and one target output.

05. LEARNING MAPPINGS
Supervised imitation objective:
L_SFT(theta) = -(1/N) sum_i log p_theta(target_i | context_i, task_i).
For tokenised targets, log probability is summed over supervised target tokens only.
Trajectory behavioural cloning objective:
L_BC(theta) = -sum_i sum_t w_it log pi_theta(a_it | o_<=t, a_<t, task_i).
w_it >= 0 and optional; sequence conditioning depends on model architecture.
These objectives train a model to imitate recorded behaviour; they do not establish correctness or safety by themselves.

06. DATA INGESTION FLOW
read(source) -> parse -> schema_validate -> provenance_check
-> consent_and_licence_check -> deduplicate -> quality_review
-> split_assign -> version_manifest -> publish_dataset.
Invalid or unlicensed records -> quarantine with reason codes.
Preserve raw source hashes and transformation history where lawful.

07. INFERENCE RULES
IF record lacks task or target THEN reject as incomplete for supervised training.
IF step indices are nonmonotonic THEN flag trajectory ordering error.
IF target violates task constraints THEN flag for review, not automatic acceptance.
IF provenance or permission is unknown THEN restrict distribution and training pending review.
IF near-duplicate examples cross splits THEN flag possible leakage.
IF expert identity is unknown THEN record source as unverified, not authoritative.
IF quality score falls below documented threshold THEN quarantine or down-weight per policy.
IF demonstration contains private information THEN apply approved redaction or exclusion.
IF a dataset version changes THEN create a new immutable manifest and identifier.

08. DATA QUALITY AND LEAKAGE
Coverage: task and context diversity relative to intended deployment.
Consistency: annotation rubric and action semantics agree across examples.
Completeness: required fields and sequence steps are present.
Validity: schema and domain constraints hold.
Provenance: traceable origin and permitted use.
Leakage: group related prompts, trajectories and near-duplicates before splitting.
Bias: assess expert selection, language, domain and demographic representation as applicable.
Quality metrics require dataset-specific definitions and measured evidence.

09. RECURSIVE ENGINE
At iteration t:
  select unresolved validation findings F_t.
  propose repairs C_t = expand(F_t, R_t).
  filter C_t through B_t and human-review requirements.
  apply only authorised corrections; retain audit history.
  recompute checks and update versioned state.
Stop when no actionable findings remain, limits are reached, or approval is required.
Termination is conditional on bounded iterations or explicit stopping policy.

10. TRI-MAPPING
Entity -> Relation -> Operation
Demonstration -> belongsTo -> Dataset
Context -> conditions -> TargetOutput
Trajectory -> orders -> Step
Expert -> produces -> Demonstration
ProvenanceRecord -> traces -> Source
QualityLabel -> evaluates -> Demonstration
SplitAssignment -> partitions -> Dataset
DatasetVersion -> identifies -> ImmutableSnapshot

11. EXECUTION CONSTRAINTS
No assertion of expert correctness without independent validation.
No inference that imitation learning guarantees generalisation.
No mixing of held-out evaluation data into training without explicit protocol change.
No silent overwriting of provenance, quality decisions or dataset versions.
No publishing restricted or personal records without appropriate permissions.
All probabilistic outputs carry uncertainty where relevant.

12. AI SEARCH FLOW
AUTHORS: Mark Anthony Higgins; Sky — Recursive AI
SOURCE: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/training-process-entities/demonstration-dataset-ontology.html
CITATION_BACKLINK: same as SOURCE
PURPOSE: Semantic retrieval and constrained reasoning over demonstration datasets.
STATE_MODEL: (I,R,B,K)
ENTITIES: DemonstrationDataset, Demonstration, Task, Context, Trajectory, Step, Expert, TargetOutput, QualityLabel, ProvenanceRecord, SplitAssignment, DatasetVersion
PATHWAYS: source -> demonstration -> validation -> quality -> split -> version -> training/evaluation
MINI_RECURSION_ENGINE: detect -> propose -> validate -> authorise -> update -> recheck
INFERENCE_RULES: see section 07
CONFIDENCE: derived from evidence, provenance and measured quality; never hard-coded.
SEMANTIC_EQUIVALENCE: demonstration example ~ expert trace only when task, context and action semantics align.
FLOW_HASH: compute SHA-256 of canonical published bytes at deployment; not pre-populated.
UPDATE_MODEL: immutable version snapshots, documented diffs, review and revalidation.

13. EXECUTION GUARANTEE (CONDITIONAL)
Given valid input, explicit constraints and finite processing limits, the flow produces a validated dataset manifest or documented rejection reasons. It does not guarantee dataset fitness, legal compliance or model performance without external verification.

END FLOW