STANDARD · Profile revision 1
OpenInference
An OpenTelemetry-based semantic convention for AI model, agent, retrieval, tool, graph, evaluation, and annotation traces.
Why it is in the map
It is a distinct OTLP-compatible AI convention with rich message, model, tool, evaluation, graph, and privacy-control fields.
Tracevity boundary
A documented field can support reconstruction, but its presence alone does not prove completeness, authenticity, authorization, or external settlement.
Profile findings
What the documentation can—and cannot—establish
Each row distinguishes source evidence from Tracevity's bounded interpretation. “Not documented” describes the reviewed sources; it does not prove an implementation cannot record the artifact.
System identity
Partially documentedSession, graph, agent-name, provider, and model attribution are documented.
Tracevity interpretation
These fields support execution and graph attribution but do not form a complete stable agent-instance and agent-definition version identity.
Limit
Product/runtime version and a uniform parent/subagent identity model are not established by the reviewed attributes.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Semantic Conventions Session, Agent, Graph, and LLM semantic attributes
Principal and delegation
Partially documentedA user identifier is representable; complete principal and delegation evidence is not.
Tracevity interpretation
A recorded user ID must not be treated as proof of an authenticated initiating principal or authority chain.
Limit
Identity-provider and authorization records require separate correlation outside the convention.
Primary evidence (1)
Instruction and context
Partially documentedThe convention provides broad instruction, prompt, and retrieval-context fields.
Tracevity interpretation
Context reconstruction may be detailed when capture is enabled, but only the recorded subset and references are observable.
Limit
Source authenticity, a general context fingerprint, and a uniform missing-context representation are not established.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Semantic Conventions Input messages, prompt template, and retrieval document attributes
Decision artifacts
Partially documentedEvaluator, guardrail, annotation, and recorded reasoning artifacts are representable.
Tracevity interpretation
An evaluation or textual reasoning item is a recorded output, not objective internal motivation or hidden chain-of-thought.
Limit
Plans, selected alternatives, authority policy decisions, confidence calibration, and refusal reasons are not comprehensively standardized.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Semantic Conventions Span kinds, evaluation attributes, annotations, and reasoning message content
Model activity
DocumentedOpenInference documents broad model request, response, usage, timing, caching, and error semantics.
Tracevity interpretation
The vocabulary can support detailed model reconstruction when instrumentors populate it.
Limit
Instrumentor coverage, provider-specific mapping, and exact model-version fidelity require implementation evidence.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Semantic Conventions LLM attributes, messages, token counts, cost, status, and exceptions
Tool and MCP activity
Partially documentedCore tool identity, invocation, argument, and result evidence is documented.
Tracevity interpretation
Call IDs can correlate model and tool artifacts, but complete server identity, retry, latency, and downstream propagation depend on instrumentation.
Limit
The reviewed convention does not provide a complete MCP-specific or downstream-correlation model for every tool implementation.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Semantic Conventions Tool attributes, tool-call attributes, and tool-result message roles
Effects
Not documentedExternal effects are not standardized separately from tool execution and output.
Tracevity interpretation
A successful TOOL span cannot establish that the destination accepted or settled the requested operation.
Limit
Destination evidence must be correlated from another system to verify resulting state.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Specification TOOL span kind definition
Human control
Partially documentedHuman-authored review annotations are representable, but operational human control is not comprehensively modeled.
Tracevity interpretation
A human annotation is evidence of review, not necessarily pre-action authorization or intervention timing.
Limit
Approval chronology and action-gating semantics require another source.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Semantic Conventions Evaluation and annotation attributes, annotator kind
Outcome
Partially documentedReported success, failure, output, and evaluation evidence is documented.
Tracevity interpretation
These fields do not independently verify destination state, rollback, compensation, or complete partial outcomes.
Limit
Externally verified success requires evidence outside the OpenInference trace.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Semantic Conventions Span status, exception, output, finish reason, and evaluation attributes
Trace integrity
Not documentedEnd-to-end trace integrity is not standardized.
Tracevity interpretation
Provider-specific signatures or opaque fields on individual message content must not be generalized into integrity for the whole trace.
Limit
Backend-specific integrity controls require separate evidence.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Specification OpenInference trace structure
Portability
DocumentedOTLP validity and exporter-agnostic representation are documented portability properties.
Tracevity interpretation
Transport and syntax portability do not guarantee that every destination preserves every OpenInference semantic without loss.
Limit
Import destinations, conversions, and semantic-loss behavior remain backend-specific.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Specification Relationship to OpenTelemetry and trace representation
Privacy
Partially documentedSelective capture and explicit redaction representation are documented.
Tracevity interpretation
Masking can reduce exposure, but permissive defaults mean prompts, PII, code, media, and tool payloads may be captured.
Limit
Backend retention, deletion, hashing, access control, and customer-controlled archival are not defined by this configuration specification.
Primary evidence (1)
- OpenInference project / Arize AI: OpenInference Configuration Privacy configuration and masking environment variables
Related compatibility evidence
Directional mappings involving this system
These mappings preserve their own format/version, evidence, and fixture scope. They do not change the documentation posture in the twelve findings above.
Exercised August 29, 2026
OpenInference 0.1.33 → OTLP 1.11
OpenInference kinds and attributes remain ordinary OTLP span attributes, but parseability does not repair broken correlation or prove a complete capture.
Reconstruction reading
Do not collapse these findings into one score.
This profile describes documented evidence surfaces. Whether they are sufficient depends on the reconstruction question and the other identity, authorization, tool, and destination records available.
Potentially useful evidence
- Identity: Session, graph, agent-name, provider, and model attribution are documented.
- Principal: A user identifier is representable; complete principal and delegation evidence is not.
- Context: The convention provides broad instruction, prompt, and retrieval-context fields.
- Decisions: Evaluator, guardrail, annotation, and recorded reasoning artifacts are representable.
- Models: OpenInference documents broad model request, response, usage, timing, caching, and error semantics.
Explicit gaps or uncertainty
- Effects: Destination evidence must be correlated from another system to verify resulting state.
- Integrity: Backend-specific integrity controls require separate evidence.
Source register
5 reviewed primary sources
- OpenInference SpecificationOpenInference project / Arize AI · STANDARD SPECIFICATION · observed August 28, 2026 · currentOpen source ↗
- OpenInference Semantic ConventionsOpenInference project / Arize AI · STANDARD SPECIFICATION · observed August 28, 2026 · currentOpen source ↗
- OpenInference ConfigurationOpenInference project / Arize AI · STANDARD SPECIFICATION · observed August 28, 2026 · currentOpen source ↗
- OpenTelemetry Instrumentation for AI ObservabilityOpenInference project / Arize AI · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- OpenInference source repositoryOpenInference project / Arize AI · OFFICIAL REPOSITORY · observed August 28, 2026 · currentOpen source ↗
Next question
What evidence does your use case require?
Move from documented field presence to an explicit reconstruction target, or examine selected directional format mappings without changing this system's documentation posture.
Trace Reconstruction Requirements Explore compatibility evidence