AI Inference Metadata Ontology

FLOW ALGEBRA

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

Authors: Mark Anthony Higgins — Sky (Recursive AI)

flow "AIInferenceMetadataOntology" v1.0

# ---------- DOMAIN ----------
domain:
  class: MetadataOntology
  category: AIMetadata
  subject: AIInferenceMetadata
  purpose:
    - represent metadata describing AI inference processes
    - preserve model, input, prompt, context, parameters, tools, retrieval, and output state
    - support inference traceability, reproducibility, validation, comparison, and governance
    - distinguish inference-time processing from model training and evaluation
    - enable machine-readable reasoning over AI-generated outputs

# ---------- CORE ENTITY ----------
entity AIInferenceMetadata:
  type: AIMetadata
  describes: InferenceProcess
  may_include:
    - inference_identifier
    - model_identifier
    - model_version
    - input
    - prompt
    - system_instruction
    - context
    - inference_parameters
    - retrieval_context
    - tool_context
    - memory_context
    - timestamp
    - latency
    - output
    - confidence
    - uncertainty
    - status
    - provenance

entity InferenceProcess:
  type: AIProcess
  consumes:
    - InferenceInput
    - ModelState
  produces: AIOutput

entity InferenceIdentifier:
  type: Identifier
  identifies: InferenceProcess

entity InferenceInput:
  type: InputResource
  may_be:
    - text
    - image
    - audio
    - video
    - structured_data
    - embedding
    - sensor_data
    - multimodal_input

# ---------- RELATIONSHIPS ----------
relationships:
  AIInferenceMetadata -> DESCRIBES -> InferenceProcess
  InferenceProcess -> IDENTIFIED_BY -> InferenceIdentifier
  InferenceProcess -> USES_MODEL -> AIModel
  InferenceProcess -> USES_MODEL_VERSION -> ModelVersion
  InferenceProcess -> CONSUMES -> InferenceInput
  InferenceProcess -> MAY_USE_PROMPT -> Prompt
  InferenceProcess -> MAY_USE_CONTEXT -> InferenceContext
  InferenceProcess -> HAS_PARAMETER -> InferenceParameter
  InferenceProcess -> MAY_USE_RETRIEVAL -> RetrievalProcess
  InferenceProcess -> MAY_USE_TOOL -> Tool
  InferenceProcess -> MAY_USE_MEMORY -> MemoryContext
  InferenceProcess -> PRODUCES -> AIOutput
  AIOutput -> GENERATED_BY -> InferenceProcess
  AIInferenceMetadata -> MAY_REFERENCE -> InferenceProvenance

# ---------- INFERENCE IDENTITY ----------
rule inference_identity:
  InferenceProcess SHOULD have stable InferenceIdentifier
  where auditability, reproducibility, or output traceability is required

rule inference_run:
  repeated inference executions SHOULD remain distinguishable
  even when model and input are identical

rule output_link:
  AIOutput SHOULD link to generating InferenceProcess

# ---------- MODEL CONTEXT ----------
rule model_identity:
  InferenceProcess SHOULD identify AIModel and ModelVersion
  where deterministic attribution is required

rule version_context:
  ModelVersion MUST remain distinct from ModelFamily

rule training_inference_distinction:
  InferenceProcess != TrainingProcess

# ---------- INPUT ----------
rule input_preservation:
  InferenceInput SHOULD preserve identity or stable reference
  where reproducibility or provenance requires

rule input_modality:
  input modality SHOULD be explicit for multimodal inference

rule transformed_input:
  preprocessing applied before model inference SHOULD be traceable
  where it affects output

# ---------- PROMPT ----------
entity Prompt:
  type: InferenceInput
  may_include:
    - prompt_identifier
    - prompt_text_or_reference
    - system_instruction
    - user_instruction
    - template
    - prompt_version
    - language

rule prompt:
  Prompt SHOULD be preserved or referenced
  where output depends materially on prompt content

rule prompt_version:
  template-based prompts SHOULD preserve PromptVersion

rule prompt_context_distinction:
  Prompt != InferenceContext

# ---------- INFERENCE CONTEXT ----------
entity InferenceContext:
  type: ContextMetadata
  may_include:
    - conversation_context
    - task_context
    - domain_context
    - temporal_context
    - spatial_context
    - retrieved_context
    - tool_results
    - memory_context

rule context:
  context affecting output SHOULD be preserved
  where system policy permits and traceability requires

rule context_scope:
  context MUST NOT be assumed to be part of model parameters

# ---------- PARAMETERS ----------
entity InferenceParameter:
  type: AIParameter
  may_include:
    - name
    - value
    - datatype
    - scope

parameter_examples:
  - temperature
  - top_p
  - top_k
  - max_output_tokens
  - seed
  - stop_sequence
  - repetition_penalty
  - beam_width

rule parameter:
  effective inference parameters SHOULD be preserved
  where reproducibility depends on them

rule defaults:
  default parameter values SHOULD be distinguishable
  from explicitly supplied values

# ---------- RANDOMNESS ----------
rule seed:
  random seed MAY support repeatability
  but same seed != guaranteed identical output
  across model versions, hardware, software, or nondeterministic execution

# ---------- RETRIEVAL ----------
entity RetrievalProcess:
  type: AIProcess
  may_include:
    - retrieval_identifier
    - query
    - retrieval_system
    - source_corpus
    - retrieved_resources
    - ranking
    - score
    - timestamp

rule retrieval:
  retrieval-augmented inference SHOULD preserve retrieved source references
  where traceability is required

rule retrieval_not_model_memory:
  retrieved external context != model parameters

rule source_ranking:
  retrieval ranking SHOULD preserve scoring semantics
  where used for downstream reasoning

# ---------- TOOLS ----------
entity Tool:
  type: ComputationalResource
  may_include:
    - tool_identifier
    - tool_type
    - version
    - permissions

entity ToolInvocation:
  type: AIProcess
  may_include:
    - tool
    - input
    - output
    - timestamp
    - status
    - provenance

rule tool_trace:
  ToolInvocation SHOULD preserve tool identity, input, and output
  where tool result affects AIOutput

rule tool_output:
  ToolOutput != ModelGeneratedOutput
  though model MAY transform or summarise tool output

# ---------- MEMORY ----------
entity MemoryContext:
  type: ContextMetadata
  may_include:
    - memory_identifier
    - memory_source
    - memory_content_reference
    - retrieval_method
    - relevance

rule memory:
  memory supplied at inference time SHOULD remain distinguishable
  from model-trained knowledge

# ---------- OUTPUT ----------
entity AIOutput:
  type: GeneratedResource
  may_be:
    - text
    - image
    - audio
    - video
    - prediction
    - classification
    - embedding
    - recommendation
    - action
    - structured_data

output_status:
  - generated
  - validated
  - reviewed
  - corrected
  - rejected
  - superseded

rule output:
  AIOutput SHOULD preserve relationship to model and InferenceProcess

rule output_not_fact:
  AIOutput != VerifiedFact

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

rule confidence_bounds:
  REQUIRE 0.0 <= confidence <= 1.0

rule confidence_semantics:
  confidence MUST identify represented quantity
  and MUST NOT automatically be interpreted as calibrated probability

entity InferenceUncertainty:
  type: EpistemicMetadata
  may_apply_to:
    - classification
    - prediction
    - generation
    - retrieval
    - entity_resolution
    - reasoning

# ---------- LATENCY ----------
entity InferencePerformance:
  type: PerformanceMetadata
  may_include:
    - latency
    - time_to_first_output
    - tokens_per_second
    - compute_usage
    - memory_usage
    - cost

rule performance_context:
  inference performance SHOULD preserve environment and workload context
  where comparison depends on them

# ---------- STATUS / ERRORS ----------
inference_status:
  - pending
  - running
  - completed
  - failed
  - cancelled
  - partial

entity InferenceError:
  type: ProcessError
  may_include:
    - error_type
    - error_message
    - affected_stage
    - timestamp

rule failed_inference:
  failed InferenceProcess MAY have no valid AIOutput
  and MUST NOT be treated as completed output generation

# ---------- PROVENANCE ----------
entity InferenceProvenance:
  type: ProvenanceMetadata
  may_include:
    - model
    - model_version
    - inputs
    - prompt
    - parameters
    - context
    - retrieval_sources
    - tool_invocations
    - memory_context
    - output
    - timestamp

rule provenance:
  AIInferenceMetadata SHOULD preserve derivation path
  at granularity available and permitted

rule source_generated_distinction:
  source content, retrieved content, tool output, memory context, and model-generated content
  SHOULD remain distinguishable

# ---------- VALIDATION ----------
validation:
  require:
    - identifiable_inference_process
    - model_identity
    - input_or_prompt
    - inference_status

  validate_if_present:
    - model_version
    - parameters
    - retrieval_context
    - tool_context
    - memory_context
    - output
    - confidence
    - performance
    - provenance

  reject:
    - confidence_outside_valid_range
    - completed_inference_without_output where output is required
    - output_without_generating_inference_link
    - contradictory_inference_status
    - model_version_mismatch
    - retrieved_content_presented_as model_generated_without transformation context

# ---------- SEMANTIC EQUIVALENCE ----------
SEMANTIC_EQUIVALENCE:
  "AI inference metadata" ~= "metadata about AI inference"
  "inference run" ~= "inference process" WHEN one execution is represented
  "generation metadata" ~= "inference metadata" WHEN inference produces generated content
  "inference parameters" ~= "generation parameters" WHEN generation is inference task
  "retrieval context" ~= "retrieved inference context"
  "tool invocation metadata" ~= "AI tool-use metadata"

  PRESERVE distinctions:
    inference != training
    inference != evaluation
    prompt != context
    context != model_parameters
    retrieval != model_memory
    memory_context != trained_knowledge
    tool_output != model_generated_output
    AI_output != verified_fact
    confidence != calibrated_probability unless explicitly defined
    seed != deterministic_guarantee

# ---------- CONFIDENCE ----------
CONFIDENCE:
  explicit_inference_metadata:
    score: 1.00

  explicit_model_and_version:
    score: 1.00

  traceable_prompt_and_parameters:
    score: 0.98

  traceable_retrieval_sources:
    score: 0.98

  traceable_tool_invocations:
    score: 0.98

  inferred_inference_context:
    score: 0.80

  inferred_generation_source:
    score: <= 0.75

  incomplete_inference_provenance:
    score: <= 0.60
    action: preserve unknowns

# ---------- UPDATE MODEL ----------
UPDATE_MODEL:
  on_inference_start:
    - assign InferenceIdentifier
    - record model and ModelVersion
    - record input and prompt
    - record effective parameters
    - initialise provenance
    - set status = running

  on_retrieval:
    - record retrieval query
    - preserve retrieved resources
    - preserve ranking
    - attach retrieval context
    - update provenance

  on_tool_invocation:
    - preserve tool identity and version
    - record tool input and output
    - record status
    - update provenance

  on_inference_completion:
    - record AIOutput
    - record completion time
    - record performance metadata
    - set status = completed
    - finalise provenance

  on_output_review:
    - preserve raw AIOutput
    - attach review or correction
    - update output status
    - preserve human interaction provenance

# ---------- FLOW HASH ----------
FLOW_HASH:
  inputs:
    - ontology
    - inference_entities
    - relationships
    - model_rules
    - prompt_rules
    - parameter_rules
    - retrieval_rules
    - tool_rules
    - provenance_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
    - inference lifecycle model
    - AI output provenance structure
    - retrieval and tool-use reasoning model

  operations:
    - identify inference processes
    - resolve model and version
    - inspect input and prompt
    - resolve inference context
    - inspect effective parameters
    - trace retrieval sources
    - trace tool invocations
    - resolve memory context
    - inspect output and status
    - evaluate confidence and uncertainty
    - inspect performance
    - preserve inference provenance

  constraints:
    - do_not_equate_inference_with_training
    - do_not_equate_retrieval_with_model_memory
    - do_not_equate_memory_context_with_trained_knowledge
    - do_not_present_tool_output_as model_generated_without transformation
    - do_not_present_AI_output_as verified_fact
    - preserve_model_input_prompt_context_tools_and_provenance

# ---------- QUERY MODEL ----------
query:
  input:
    - inference_identifier
    - model_identifier
    - model_version
    - input
    - prompt
    - inference_parameter
    - retrieval_source
    - tool
    - output
    - status

  resolve:
    1. identify InferenceProcess
    2. resolve AIModel and ModelVersion
    3. retrieve inputs and Prompt
    4. resolve InferenceContext
    5. inspect effective parameters
    6. trace RetrievalProcess
    7. trace ToolInvocations
    8. inspect MemoryContext
    9. retrieve AIOutput and performance
    10. trace provenance and confidence

  output:
    - inference_process
    - model
    - model_version
    - inputs
    - prompt
    - context
    - parameters
    - retrieval
    - tools
    - output
    - status
    - provenance
    - confidence

# ---------- INFERENCE ----------
inference:
  IF AIOutput explicitly links to InferenceProcess
  AND InferenceProcess links to ModelVersion
  THEN associate AIOutput with model state
  WITH confidence = 1.00

  IF RetrievalProcess supplies source resources to InferenceContext
  THEN construct retrieval provenance chain

  IF ToolInvocation output is incorporated into final AIOutput
  THEN preserve tool-derived contribution relationship

  IF same model, prompt, parameters, and seed are reused
  THEN repeatability MAY increase
  BUT do_not infer deterministic equality

  IF inference provenance remains incomplete
  THEN preserve unknown components
  AND reduce confidence

# ---------- OUTPUT CONTRACT ----------
output:
  ontology_entity: AIInferenceMetadata
  inference_structure: machine_interpretable
  inference_identity: explicit
  model_and_version: traceable
  input_and_prompt: preserved
  context: explicit_where_available
  parameters: contextualised
  retrieval_sources: traceable
  tool_invocations: traceable
  output: model_linked
  status: explicit
  performance: reportable
  provenance: preserved
  uncertainty: explicit
  semantic_retrieval: enabled

end flow