OBSERVABILITY PLATFORM · Profile revision 1
Arize Phoenix
An open-source observability and evaluation platform that accepts OTLP traces and uses OpenInference semantics for model, retrieval, agent, and tool activity.
Why it is in the map
Phoenix is a high-signal OpenInference/OTLP implementation with sessions, annotations, export/query, self-hosted privacy, and retention controls.
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
DocumentedTrace/span hierarchy, session identity, timing, and resource attributes support distributed execution correlation.
Tracevity interpretation
Stable agent definition and version still depend on the emitting instrumentor.
Limit
Canonical agent instance and agent-definition/version fields are not guaranteed for every trace.
Primary evidence (1)
- Arize AI: What are Traces? Span context and hierarchy
Principal and delegation
Partially documentedApplication-supplied user attribution can be recorded.
Tracevity interpretation
A user attribute is not verified initiating-principal or authority evidence.
Limit
Authentication, delegation chain, credential type/scope/expiry, and elevation are not canonical.
Primary evidence (1)
- Arize AI: Sessions Session and user attributes
Instruction and context
DocumentedInstructions, prompt data, retrieval content, and contextual metadata are representable.
Tracevity interpretation
Capture does not prove source authenticity, completeness, or consistent versioning.
Limit
Context fingerprints and explicit missing-context representation are not standardized across instrumentors.
Primary evidence (1)
- Arize AI: What are Traces? OpenInference span kinds and attributes
Decision artifacts
Partially documentedEvaluator output and explicit assessment artifacts can be retained with provenance category.
Tracevity interpretation
An explanation is a recorded assessment, not hidden reasoning or proof of the agent's motive.
Limit
Plan, decision summary, selected alternative, policy decision, calibrated confidence, and refusal are not universal.
Primary evidence (1)
- Arize AI: Annotations API Annotation data and annotator kind
Model activity
DocumentedCore model interaction and performance evidence is representable.
Tracevity interpretation
Exact provider version and field completeness depend on the chosen instrumentor and capture settings.
Limit
Provider build/version, raw request/response, cache, and token details are not guaranteed across integrations.
Primary evidence (1)
- Arize AI: What are Traces? OpenInference LLM span kind
Tool and MCP activity
DocumentedTool invocation, result, error, timing, and trace correlation are representable.
Tracevity interpretation
Tool telemetry does not establish authorization or downstream settlement.
Limit
MCP server, call ID, retry, credential authority, and downstream correlation depend on instrumentation.
Primary evidence (1)
- Arize AI: What are Traces? OpenInference TOOL span kind
Effects
Not documentedPhoenix represents tool activity, not a universal external-effect settlement chain.
Tracevity interpretation
A successful tool span cannot establish that the destination changed.
Limit
Requested effect, acceptance, external effect state, and state delta are not canonical Phoenix records.
Primary evidence (1)
- Arize AI: What are Traces? TOOL span semantics
Human control
Partially documentedHuman review and evaluation evidence can be recorded.
Tracevity interpretation
Human annotation is not necessarily pre-action approval, rejection, or intervention.
Limit
Approval, rejection, edit, interruption, takeover, rollback request, authority, and timing are not universal.
Primary evidence (1)
- Arize AI: Annotations API HUMAN annotator kind
Outcome
Partially documentedExecution-reported and evaluated outcomes can be correlated in a trace.
Tracevity interpretation
Span success and evaluation do not independently verify external state.
Limit
Externally verified success, settlement ambiguity, rollback, and compensation are not canonical.
Primary evidence (1)
- Arize AI: What are Traces? Span status, outputs, and errors
Trace integrity
Not documentedTimestamps and retained hierarchy support correlation, not independently verifiable integrity.
Tracevity interpretation
Self-hosted custody does not itself make a record immutable.
Limit
Clock provenance, immutability, append-only enforcement, signing, tamper evidence, and independent verification are absent.
Primary evidence (1)
- Arize AI: Data retention Retention defaults and deletion
Portability
DocumentedOTLP, OpenInference, and query/export surfaces provide strong transport and analysis portability.
Tracevity interpretation
Phoenix annotations and vendor extensions may not round-trip through every OTel destination.
Limit
Schema versions, import/export symmetry, documented conversions, and exact semantic loss remain incomplete.
Primary evidence (1)
- Arize AI: How does tracing work? OTLP collector and OpenInference instrumentors
Privacy
DocumentedSelf-hosted custody and retention/deletion controls are documented while rich trace payloads may remain sensitive.
Tracevity interpretation
No-egress custody is not automatic redaction, filtering, minimization, or safe retention.
Limit
Client-side redaction, hashing, screenshot controls, and actual capture completeness depend on instrumentation and deployment.
Primary evidence (1)
- Arize AI: Privacy Self-hosted data egress
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: Trace/span hierarchy, session identity, timing, and resource attributes support distributed execution correlation.
- Principal: Application-supplied user attribution can be recorded.
- Context: Instructions, prompt data, retrieval content, and contextual metadata are representable.
- Decisions: Evaluator output and explicit assessment artifacts can be retained with provenance category.
- Models: Core model interaction and performance evidence is representable.
Explicit gaps or uncertainty
- Effects: Requested effect, acceptance, external effect state, and state delta are not canonical Phoenix records.
- Integrity: Clock provenance, immutability, append-only enforcement, signing, tamper evidence, and independent verification are absent.
Source register
7 reviewed primary sources
- What are Traces?Arize AI · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- How does tracing work?Arize AI · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Extract data from spansArize AI · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- SessionsArize AI · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Annotations APIArize AI · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- PrivacyArize AI · OFFICIAL POLICY · observed August 28, 2026 · currentOpen source ↗
- Data retentionArize AI · OFFICIAL POLICY · 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