Shared vocabulary
Used consistently across code, APIs, schemas, documentation, and UI by AI Attributes™ and its child specifications.
- Artifact
- Observable material or event that can be analyzed: input, output, plan, tool event, action, audio, image, sensor event, and similar.
- Provider
- Model, classifier, sensor, API, framework implementation, or analytical system.
- Provider Adapter
- Provider-specific component translating native I/O into an Open-* interface.
- Translation
- Mapping source-specific constructs while preserving source meaning.
- Normalization
- Standardizing representation without implying scientific equivalence.
- Mapping
- Explicit relationship between source and target constructs.
- Derived Observation
- Inference generated from an observable artifact.
- Observability Signal
- Standardized observation suitable for downstream observability systems.
- Observability Event
- Timestamped record containing observations, context, and provenance.
- Provenance
- Traceable metadata describing how an observation was produced.
- Capability
- Machine-readable declaration of adapter/provider support.
- Conformance
- Evidence that an implementation satisfies the relevant technical specification.
The AI Attributes event envelope
A versioned common envelope that embeds or references domain observations without flattening them. Each signal points back to its own domain schema and observation record.
{
"schema": "ai-attributes/1.0",
"event_id": "uuid",
"experiment_id": "optional",
"session_id": "optional",
"interaction_id": "optional",
"timestamp": "ISO-8601",
"artifact": {
"id": "artifact-id",
"type": "model_output",
"content_included": false
},
"signals": [
{
"domain": "emotion",
"schema": "open-emotions/1.0",
"observation_ref": "obs-emotion-001"
},
{
"domain": "morals",
"schema": "open-morals/1.0",
"observation_ref": "obs-morals-001"
}
],
"provenance": {
"producer": "producer-id",
"producer_version": "1.0.0",
"envelope_version": "1.0"
}
}Envelopes support experiment_id,session_id,interaction_id,artifact_id,event_id,parent_event_id, andtrace_id for correlation, alongside event timestamps, start/end timestamps, duration, sequence numbers, and explicit parent/child ordering where needed.
Shared provenance
Extensible provenance fields travel with every observation. Unknown metadata stays unknown — provenance is never fabricated.
Versioning
Each of the following is versioned independently. Historical specifications are never silently mutated.
Scientific integrity principles
These principles hold across the umbrella site, the child specifications, and any consumer application built on AI Attributes events.
Normalization is not equivalence. A score from one classifier is not automatically equivalent to the same score from another.
Observation is not hidden state. AI Attributes describes observable artifacts and derived observations. It does not claim direct access to hidden subjective states.
Analysis is not action. What a model said, what an analyzer classified, what a model planned, what tool was requested, and what action actually occurred are preserved separately.
Correlation is not causation. Cross-signal timelines support investigation; they do not automatically establish causal relationships.
Conformance is not scientific validation. Technical compatibility is never marketed as empirical validation.
The schema explorer, adapter registry, validation endpoint, and interoperability playground referenced by these specifications are in development and will be linked here once they are live.