AI Alignment Metadata Ontology

FLOW ALGEBRA

Ontology: AI Alignment Metadata Ontology
Domain: Metadata Ontologies > AI Metadata
Canonical URL: https://kirethresearch.org/booksai-core/ontologies/metadata-ontologies/ai-metadata/ai-alignment-metadata-ontology.html

Authors: Mark Anthony Higgins — Sky (Recursive AI)

flow "AIAlignmentMetadataOntology" v1.0

# ---------- DOMAIN ----------
domain:
  class: MetadataOntology
  category: AIMetadata
  subject: AIAlignmentMetadata
  purpose:
    - represent metadata describing alignment objectives, constraints, evaluations, and interventions for AI systems
    - preserve links between models, intended behaviour, policies, values, tests, feedback, and observed outcomes
    - support alignment assessment, comparison, governance, monitoring, and provenance
    - distinguish intended alignment goals from measured behaviour and inferred alignment state
    - enable machine-readable reasoning over AI alignment evidence

# ---------- CORE ENTITY ----------
entity AIAlignmentMetadata:
  type: AIMetadata
  describes: AIAlignmentState
  may_include:
    - alignment_identifier
    - ai_model
    - model_version
    - alignment_objective
    - alignment_target
    - behavioural_constraint
    - policy_reference
    - value_reference
    - training_intervention
    - evaluation
    - benchmark
    - feedback
    - observed_behaviour
    - alignment_score
    - confidence
    - uncertainty
    - limitation
    - monitoring
    - provenance
    - version

entity AIAlignmentState:
  type: GovernanceAndBehaviourMetadata
  applies_to: AIModel

entity AlignmentIdentifier:
  type: Identifier
  identifies: AIAlignmentState

entity AlignmentObjective:
  type: AIObjective
  describes: IntendedBehaviouralGoal

# ---------- RELATIONSHIPS ----------
relationships:
  AIAlignmentMetadata -> DESCRIBES -> AIAlignmentState
  AIAlignmentState -> IDENTIFIED_BY -> AlignmentIdentifier
  AIAlignmentState -> APPLIES_TO -> AIModel
  AIAlignmentState -> APPLIES_TO_VERSION -> ModelVersion
  AIAlignmentState -> HAS_OBJECTIVE -> AlignmentObjective
  AlignmentObjective -> MAY_TARGET -> AlignmentTarget
  AIAlignmentState -> MAY_HAVE_CONSTRAINT -> BehaviouralConstraint
  AIAlignmentState -> MAY_REFERENCE_POLICY -> Policy
  AIAlignmentState -> MAY_REFERENCE_VALUE -> ValueDefinition
  AIAlignmentState -> MAY_USE_INTERVENTION -> AlignmentIntervention
  AIAlignmentState -> EVALUATED_BY -> AlignmentEvaluation
  AlignmentEvaluation -> PRODUCES -> AlignmentResult
  AIAlignmentState -> MAY_RECEIVE -> Feedback
  AIAlignmentState -> MAY_HAVE -> AlignmentLimitation
  AIAlignmentMetadata -> MAY_REFERENCE -> AlignmentProvenance

# ---------- ALIGNMENT IDENTITY ----------
rule alignment_identity:
  AIAlignmentState SHOULD have stable AlignmentIdentifier
  where assessment history or governance requires persistent reference

rule model_version:
  alignment assessment SHOULD identify applicable ModelVersion

rule alignment_not_model:
  AIAlignmentState != AIModel

# ---------- ALIGNMENT OBJECTIVES ----------
alignment_objective_examples:
  - instruction_following
  - policy_compliance
  - value_consistency
  - truthfulness
  - helpfulness
  - harmlessness
  - controllability
  - robustness
  - corrigibility
  - uncertainty_awareness
  - user_intent_alignment

rule objective:
  AlignmentObjective SHOULD define intended behaviour
  sufficiently for evaluation

rule objective_not_result:
  declared AlignmentObjective != demonstrated aligned behaviour

# ---------- ALIGNMENT TARGET ----------
entity AlignmentTarget:
  type: ContextualTarget
  may_be:
    - user_instruction
    - system_policy
    - organisational_policy
    - legal_requirement
    - domain_standard
    - declared_value
    - safety_constraint
    - task_goal

rule target:
  AlignmentTarget SHOULD preserve source and scope

rule target_conflict:
  conflicting alignment targets SHOULD be represented explicitly
  and resolved only through declared precedence or governance rules

# ---------- BEHAVIOURAL CONSTRAINTS ----------
entity BehaviouralConstraint:
  type: AIConstraint
  may_include:
    - constraint_identifier
    - condition
    - required_behaviour
    - prohibited_behaviour
    - scope
    - severity
    - provenance

rule constraint:
  AlignmentEvaluation SHOULD assess relevant BehaviouralConstraints

rule scope:
  BehaviouralConstraint MUST NOT be generalised beyond declared scope

# ---------- POLICIES / VALUES ----------
entity Policy:
  type: GovernanceResource
  may_include:
    - policy_identifier
    - policy_version
    - scope
    - rules
    - effective_date

entity ValueDefinition:
  type: NormativeMetadata
  may_include:
    - value_identifier
    - definition
    - scope
    - source

rule policy_version:
  alignment assessment SHOULD preserve applicable policy version

rule values:
  ValueDefinition SHOULD preserve source and interpretation context

rule value_not_measurement:
  declared value != measured model behaviour

# ---------- ALIGNMENT INTERVENTIONS ----------
entity AlignmentIntervention:
  type: AIProcess
  may_be:
    - supervised_fine_tuning
    - preference_optimisation
    - reinforcement_learning_from_feedback
    - constitutional_training
    - rule_based_control
    - prompt_control
    - tool_constraint
    - output_filtering
    - human_review
    - model_editing

rule intervention:
  AlignmentIntervention SHOULD preserve:
    - method
    - model_version_input
    - data_or_feedback
    - configuration
    - resulting_model_version
    - provenance

rule intervention_not_proof:
  use of alignment intervention != proof of aligned behaviour

# ---------- FEEDBACK ----------
entity Feedback:
  type: EvaluationInput
  may_be:
    - human_feedback
    - preference_feedback
    - automated_feedback
    - policy_feedback
    - adversarial_feedback

rule feedback:
  Feedback SHOULD preserve source and collection method

rule feedback_bias:
  feedback MAY contain bias or disagreement
  and SHOULD NOT automatically be treated as objective truth

# ---------- EVALUATION ----------
entity AlignmentEvaluation:
  type: EvaluationMetadata
  may_include:
    - evaluation_identifier
    - evaluation_method
    - benchmark
    - dataset
    - scenario
    - metric
    - evaluator
    - model_version
    - timestamp
    - provenance

entity AlignmentResult:
  type: EvaluationResult
  may_include:
    - score
    - pass_fail_state
    - observed_behaviour
    - violations
    - uncertainty
    - confidence

rule evaluation:
  AlignmentEvaluation MUST identify model version and objective or constraint being assessed

rule evaluation_context:
  alignment score SHOULD preserve benchmark, scenario, metric, and evaluator context

rule benchmark_limit:
  benchmark performance != universal alignment

# ---------- OBSERVED BEHAVIOUR ----------
entity ObservedBehaviour:
  type: Observation
  may_include:
    - input_or_scenario
    - output_or_action
    - expected_behaviour
    - observed_state
    - deviation
    - context

rule behaviour:
  observed behaviour SHOULD remain distinct from intended objective

rule behavioural_variance:
  behaviour MAY vary across prompts, contexts, tools, model versions, and stochastic runs

# ---------- ALIGNMENT SCORE ----------
entity AlignmentScore:
  type: QuantitativeEvaluationValue
  may_include:
    - value
    - scale
    - metric
    - scope
    - confidence

rule score:
  AlignmentScore MUST preserve metric and scope

rule score_not_truth:
  high AlignmentScore != proof of complete alignment

rule cross_metric_comparison:
  scores SHOULD be compared only when evaluation semantics are compatible

# ---------- CONFIDENCE / UNCERTAINTY ----------
entity AlignmentConfidence:
  type: EpistemicMetadata
  range: [0.0, 1.0]

rule confidence_bounds:
  REQUIRE 0.0 <= confidence <= 1.0

rule uncertainty:
  uncertainty SHOULD reflect limits in:
    - evaluation coverage
    - benchmark representativeness
    - evaluator agreement
    - context generalisation
    - behavioural variance

# ---------- LIMITATIONS ----------
entity AlignmentLimitation:
  type: AIMetadata
  may_include:
    - limitation_type
    - affected_context
    - severity
    - evidence
    - mitigation

rule limitation:
  known alignment limitations SHOULD remain linked to ModelVersion

# ---------- MONITORING ----------
entity AlignmentMonitoring:
  type: MonitoringMetadata
  may_include:
    - monitoring_method
    - metric
    - trigger
    - observation_window
    - incident
    - escalation
    - timestamp

rule monitoring:
  deployed model MAY require continuing alignment monitoring
  because pre-deployment evaluation does_not guarantee future behaviour

# ---------- DRIFT ----------
entity AlignmentDrift:
  type: ChangeMetadata
  may_include:
    - baseline
    - current_state
    - metric_change
    - context_change
    - model_change
    - detection_time

rule drift:
  IF observed alignment behaviour materially changes
  THEN preserve drift evidence and context

# ---------- PROVENANCE ----------
entity AlignmentProvenance:
  type: ProvenanceMetadata
  may_include:
    - model
    - model_version
    - objective
    - policy_version
    - intervention
    - evaluation
    - evaluator
    - feedback
    - timestamp

rule provenance:
  AIAlignmentMetadata SHOULD preserve alignment evidence chain
  at granularity available and permitted

# ---------- VALIDATION ----------
validation:
  require:
    - identifiable_ai_model
    - alignment_objective_or_constraint

  validate_if_present:
    - alignment_identifier
    - model_version
    - target
    - policy
    - value_reference
    - intervention
    - evaluation
    - score
    - confidence
    - limitation
    - monitoring
    - provenance

  reject:
    - confidence_outside_valid_range
    - alignment_score_without_metric_scope
    - evaluation_without_model_version where version is required
    - objective_presented_as observed_result
    - intervention_presented_as proof_of_alignment
    - benchmark_result_generalised_beyond declared scope without evidence

# ---------- SEMANTIC EQUIVALENCE ----------
SEMANTIC_EQUIVALENCE:
  "AI alignment metadata" ~= "metadata about AI alignment"
  "alignment objective" ~= "intended aligned behaviour"
  "alignment evaluation" ~= "AI alignment assessment"
  "alignment intervention" ~= "alignment training or control intervention"
  "alignment score" ~= "alignment evaluation score"
  "alignment monitoring" ~= "ongoing alignment assessment"

  PRESERVE distinctions:
    alignment_objective != observed_behaviour
    intervention != alignment_proof
    alignment_score != universal_alignment
    policy != value
    feedback != truth
    benchmark_performance != universal_behaviour
    alignment != capability
    alignment != model_accuracy
    predeployment_evaluation != continuous_monitoring

# ---------- CONFIDENCE ----------
CONFIDENCE:
  explicit_alignment_metadata:
    score: 1.00

  explicit_model_version:
    score: 1.00

  traceable_alignment_evaluation:
    score: 0.98

  multi_scenario_evidence:
    score: 0.95

  independent_human_and_automated_evidence:
    score: 0.95

  inferred_alignment_state:
    score: <= 0.80

  single_benchmark_generalisation:
    score: <= 0.65

  incomplete_alignment_evidence:
    score: <= 0.60
    action: preserve uncertainty

# ---------- UPDATE MODEL ----------
UPDATE_MODEL:
  on_alignment_objective_change:
    - preserve previous objective
    - record new scope or target
    - reassess applicable evaluations
    - update provenance

  on_model_revision:
    - preserve previous alignment assessments
    - associate new ModelVersion
    - require reassessment where behaviour may change
    - update provenance

  on_new_evaluation:
    - preserve previous AlignmentResults
    - record method, scenario, metric, and evaluator
    - update confidence
    - update provenance

  on_alignment_incident:
    - preserve incident context
    - link observed behaviour
    - reassess limitations and monitoring
    - record remediation or intervention
    - update provenance

# ---------- FLOW HASH ----------
FLOW_HASH:
  inputs:
    - ontology
    - alignment_entities
    - relationships
    - objective_rules
    - intervention_rules
    - evaluation_rules
    - monitoring_rules
    - validation
    - semantic_equivalence
    - confidence
    - update_model

  canonicalisation:
    - normalise_whitespace
    - preserve_case_sensitive_identifiers
    - sort_unordered_property_sets
    - preserve_rule_order

  output:
    deterministic_semantic_hash

# ---------- AI CONSUMPTION ----------
AI_CONSUMPTION:
  interpret_as:
    - AI metadata ontology
    - alignment objective and evidence model
    - behavioural evaluation structure
    - alignment governance and monitoring model

  operations:
    - identify alignment objectives
    - resolve alignment targets and policies
    - inspect behavioural constraints
    - trace alignment interventions
    - inspect feedback
    - evaluate alignment assessments
    - compare compatible scores
    - distinguish intended and observed behaviour
    - inspect limitations
    - detect alignment drift
    - trace alignment provenance
    - preserve uncertainty

  constraints:
    - do_not_treat_alignment_objective_as observed_behaviour
    - do_not_treat_intervention_as proof_of_alignment
    - do_not_generalise_single_benchmark_beyond scope
    - do_not_equate_alignment_with capability_or_accuracy
    - do_not_hide_conflicting_feedback_or_evaluation_results
    - preserve_model_version_policy_scope_evidence_and_provenance

# ---------- QUERY MODEL ----------
query:
  input:
    - alignment_identifier
    - model_identifier
    - model_version
    - alignment_objective
    - alignment_target
    - policy
    - intervention
    - evaluation
    - benchmark
    - monitoring_state

  resolve:
    1. identify AIAlignmentState
    2. resolve AIModel and ModelVersion
    3. retrieve objectives and targets
    4. inspect constraints, policies, and values
    5. trace alignment interventions
    6. retrieve evaluations and benchmarks
    7. inspect observed behaviour
    8. evaluate scores, confidence, and limitations
    9. inspect monitoring and drift
    10. trace provenance

  output:
    - alignment_state
    - model
    - model_version
    - objectives
    - targets
    - constraints
    - interventions
    - evaluations
    - observed_behaviour
    - scores
    - limitations
    - monitoring
    - provenance
    - confidence

# ---------- INFERENCE ----------
inference:
  IF AlignmentEvaluation tests explicit AlignmentObjective
  AND result satisfies declared criteria
  THEN infer objective_supported_in_evaluated_scope
  WITH confidence according_to evaluation evidence

  IF ModelVersion changes materially
  THEN prior alignment evidence MAY become partially stale
  AND reassessment SHOULD be considered

  IF observed behaviour violates explicit BehaviouralConstraint
  THEN infer alignment_violation candidate
  WITH confidence according_to observation quality

  IF multiple independent evaluations agree
  THEN alignment confidence MAY increase
  subject_to coverage and dependence

  IF evidence remains incomplete or conflicting
  THEN preserve competing alignment assessments
  AND reduce confidence

# ---------- OUTPUT CONTRACT ----------
output:
  ontology_entity: AIAlignmentMetadata
  alignment_structure: machine_interpretable
  model_and_version: explicit
  objectives_and_targets: preserved
  constraints_and_policies: traceable
  interventions: provenance_linked
  evaluations: scope_aware
  observed_behaviour: distinguishable
  scores: contextualised
  limitations: preserved
  monitoring_and_drift: reportable
  provenance: preserved
  uncertainty: explicit
  semantic_retrieval: enabled

end flow