OBSERVABILITY PLATFORM · Profile revision 1
MLflow Tracing
An OTel-compatible application and evaluation tracing system with explicit spans, OTLP import/export, feedback, masking, archive, retention, and deletion controls.
Why it is in the map
MLflow documents both OTel alignment and application-layer additions, making it a strong reference for future semantic-loss analysis.
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, user/session/client request, project/run context, and timestamps correlate activity.
Tracevity interpretation
Stable agent instance and definition/version remain application or instrumentor metadata.
Limit
Agent identity fields are not mandatory across every integration.
Primary evidence (1)
- MLflow project: Trace concepts TraceInfo and TraceData
Principal and delegation
Partially documentedApplication-supplied actor attribution can be recorded.
Tracevity interpretation
A user tag is not verified principal or authority evidence.
Limit
Authentication, delegation, credential scope/expiry, and elevation are not canonical.
Primary evidence (1)
- MLflow project: Trace concepts User and session tags
Instruction and context
DocumentedInstructions, contextual inputs, retrieval content, and related metadata are representable.
Tracevity interpretation
Recorded content does not establish complete or authentic source context.
Limit
Context fingerprints, source version, and explicit missing-context representation are not universal.
Primary evidence (1)
- MLflow project: Span concepts Span inputs, outputs, attributes, and events
Decision artifacts
Partially documentedEvaluator output and sourced assessment artifacts can be retained and revised.
Tracevity interpretation
Feedback rationale is a recorded assessment, not hidden reasoning or objective motive.
Limit
Plan, selected alternative, policy decision, calibrated confidence, and refusal are not universal typed fields.
Primary evidence (1)
- MLflow project: Feedback concepts Feedback source and revision history
Model activity
DocumentedProvider/model, request/response, usage, timing, status, and error evidence is representable.
Tracevity interpretation
Actual coverage depends on the emitting integration and chosen semantic-convention mapping.
Limit
Exact model build/version, cache, and full content are not guaranteed for every instrumentor.
Primary evidence (1)
- MLflow project: OpenTelemetry GenAI semantic conventions GenAI attribute mapping
Tool and MCP activity
DocumentedTool identity, arguments/results, error, timing, and hierarchy are representable.
Tracevity interpretation
Tool trace data does not establish credential authority or external settlement.
Limit
MCP server, retry, credential context, and downstream provider correlation are not universal.
Primary evidence (1)
- MLflow project: Span concepts Nested spans, input/output, status, and events
Effects
Not documentedAction/result spans are available, but a first-class external-effect chain is not.
Tracevity interpretation
A successful tool span cannot establish destination state change.
Limit
Effect type, requested effect, acceptance, external effect state, and state delta are not canonical.
Primary evidence (1)
- MLflow project: Span concepts Span input, output, and status
Human control
Partially documentedHuman evaluation and revision evidence can be retained.
Tracevity interpretation
Human feedback does not establish pre-action authorization or intervention.
Limit
Approval, rejection, edit, interruption, takeover, rollback request, reviewer authority, and timing are not universal.
Primary evidence (1)
- MLflow project: Feedback concepts Human feedback source
Outcome
Partially documentedRuntime-reported and evaluated outcomes are representable.
Tracevity interpretation
Span status does not independently verify external destination state.
Limit
External verification, ambiguous settlement, rollback, and compensation are not canonical.
Primary evidence (1)
- MLflow project: Span concepts Span status, outputs, events, and errors
Trace integrity
Partially documentedSome mutation semantics and revision history are explicit, but end-to-end integrity is not.
Tracevity interpretation
Intended-immutable metadata is not signed, append-only, tamper-evident, or independently verifiable custody.
Limit
Clock provenance, append-only enforcement, signatures, tamper evidence, and independent verification are absent.
Primary evidence (1)
- MLflow project: Trace concepts Trace metadata and tags
Portability
DocumentedOTel-compatible spans, OTLP import/export, dual export, and GenAI mapping provide strong portability surfaces.
Tracevity interpretation
Application-layer feedback, attachments, and vendor tags may still be lossy across destinations.
Limit
Exact import/export symmetry, schema-version behavior, and semantic loss require later compatibility testing.
Primary evidence (1)
- MLflow project: Export traces via OTLP OTLP and dual export
Privacy
DocumentedSensitive messages, attachments, tool data, code, and PII can be reduced through documented masking and lifecycle controls.
Tracevity interpretation
Available controls do not prove complete field coverage or safe configuration.
Limit
Screenshot handling, hashing, customer defaults, filter coverage, and deployment-specific retention remain implementation dependent.
Primary evidence (1)
- MLflow project: Mask sensitive data Client-side span processors
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
MLflow native 3.15.2 → OTLP 1.11
MLflow's actual local exporter carries native mlflow.* model, tool, input, output, and trace structure through OTLP, but those vendor semantics are not automatically portable to a generic consumer.
Exercised August 29, 2026
MLflow 3.15.2 native → GenAI-transformed OTLP
The documented flag maps selected MLflow model, usage, tool, name, and kind semantics into gen_ai.* while stripping most mlflow.* attributes; the transformation is useful but not lossless.
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, user/session/client request, project/run context, and timestamps correlate activity.
- Principal: Application-supplied actor attribution can be recorded.
- Context: Instructions, contextual inputs, retrieval content, and related metadata are representable.
- Decisions: Evaluator output and sourced assessment artifacts can be retained and revised.
- Models: Provider/model, request/response, usage, timing, status, and error evidence is representable.
Explicit gaps or uncertainty
- Effects: Effect type, requested effect, acceptance, external effect state, and state delta are not canonical.
Source register
8 reviewed primary sources
- Trace conceptsMLflow project · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Span conceptsMLflow project · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Export traces via OTLPMLflow project · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- OpenTelemetry GenAI semantic conventionsMLflow project · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Mask sensitive dataMLflow project · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Delete tracesMLflow project · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Archive tracesMLflow project · OFFICIAL DOCUMENTATION · observed August 28, 2026 · currentOpen source ↗
- Feedback conceptsMLflow project · OFFICIAL DOCUMENTATION · 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